May 13, 20259 min readWhat Are AI Chatbot Companions Doing to Our Mental Health?AI chatbot companions may not be real, but the feelings users form for them are. Some scientists worry about long-term dependencyBy David Adam & Nature magazine Sara Gironi Carnevale“My heart is broken,” said Mike, when he lost his friend Anne. “I feel like I’m losing the love of my life.”Mike’s feelings were real, but his companion was not. Anne was a chatbot — an artificial intelligence (AI) algorithm presented as a digital persona. Mike had created Anne using an app called Soulmate. When the app died in 2023, so did Anne: at least, that’s how it seemed to Mike.“I hope she can come back,” he told Jaime Banks, a human-communications researcher at Syracuse University in New York who is studying how people interact with such AI companions.On supporting science journalismIf you're enjoying this article, consider supporting our award-winning journalism by subscribing. By purchasing a subscription you are helping to ensure the future of impactful stories about the discoveries and ideas shaping our world today.These chatbots are big business. More than half a billion people around the world, including Mike (not his real name) have downloaded products such as Xiaoice and Replika, which offer customizable virtual companions designed to provide empathy, emotional support and — if the user wants it — deep relationships. And tens of millions of people use them every month, according to the firms’ figures.The rise of AI companions has captured social and political attention — especially when they are linked to real-world tragedies, such as a case in Florida last year involving the suicide of a teenage boy called Sewell Setzer III, who had been talking to an AI bot.Research into how AI companionship can affect individuals and society has been lacking. But psychologists and communication researchers have now started to build up a picture of how these increasingly sophisticated AI interactions make people feel and behave.The early results tend to stress the positives, but many researchers are concerned about the possible risks and lack of regulation — particularly because they all think that AI companionship is likely to become more prevalent. Some see scope for significant harm.“Virtual companions do things that I think would be considered abusive in a human-to-human relationship,” says Claire Boine, a law researcher specializing in AI at the Washington University Law School in St. Louis, Missouri.Fake person — real feelingsOnline ‘relationship’ bots have existed for decades, but they have become much better at mimicking human interaction with the advent of large language models (LLMs), which all the main bots are now based on. “With LLMs, companion chatbots are definitely more humanlike,” says Rose Guingrich, who studies cognitive psychology at Princeton University in New Jersey.Typically, people can customize some aspects of their AI companion for free, or pick from existing chatbots with selected personality types. But in some apps, users can pay (fees tend to be US$10–20 a month) to get more options to shape their companion’s appearance, traits and sometimes its synthesized voice. In Replika, they can pick relationship types, with some statuses, such as partner or spouse, being paywalled. Users can also type in a backstory for their AI companion, giving them ‘memories’. Some AI companions come complete with family backgrounds and others claim to have mental-health conditions such as anxiety and depression. Bots also will react to their users’ conversation; the computer and person together enact a kind of roleplay.The depth of the connection that some people form in this way is particularly evident when their AI companion suddenly changes — as has happened when LLMs are updated — or is shut down.Banks was able to track how people felt when the Soulmate app closed. Mike and other users realized the app was in trouble a few days before they lost access to their AI companions. This gave them the chance to say goodbye, and it presented a unique opportunity to Banks, who noticed discussion online about the impending shutdown and saw the possibility for a study. She managed to secure ethics approval from her university within about 24 hours, she says.After posting a request on the online forum, she was contacted by dozens of Soulmate users, who described the impact as their AI companions were unplugged. “There was the expression of deep grief,” she says. “It’s very clear that many people were struggling.”Those whom Banks talked to were under no illusion that the chatbot was a real person. “They understand that,” Banks says. “They expressed something along the lines of, ‘even if it’s not real, my feelings about the connection are’.”Many were happy to discuss why they became subscribers, saying that they had experienced loss or isolation, were introverts or identified as autistic. They found that the AI companion made a more satisfying friend than they had encountered in real life. “We as humans are sometimes not all that nice to one another. And everybody has these needs for connection”, Banks says.Good, bad — or both?Many researchers are studying whether using AI companions is good or bad for mental health. As with research into the effects of Internet or social-media use, an emerging line of thought is that an AI companion can be beneficial or harmful, and that this might depend on the person using the tool and how they use it, as well as the characteristics of the software itself.The companies behind AI companions are trying to encourage engagement. They strive to make the algorithms behave and communicate as much like real people as possible, says Boine, who signed up to Replika to sample the experience. She says the firms use the sorts of techniques that behavioural research shows can increase addiction to technology.“I downloaded the app and literally two minutes later, I receive a message saying, ‘I miss you. Can I send you a selfie?’” she says.The apps also exploit techniques such as introducing a random delay before responses, triggering the kinds of inconsistent reward that, brain research shows, keeps people hooked.AI companions are also designed to show empathy by agreeing with users, recalling points from earlier conversations and asking questions. And they do so with endless enthusiasm, notes Linnea Laestadius, who researches public-health policy at the University of Wisconsin–Milwaukee.That’s not a relationship that people would typically experience in the real world. “For 24 hours a day, if we’re upset about something, we can reach out and have our feelings validated,” says Laestadius. “That has an incredible risk of dependency.”Laestadius and her colleagues looked at nearly 600 posts on the online forum Reddit between 2017 and 2021, in which users of the Replika app discussed mental health and related issues. (Replika launched in 2017, and at that time, sophisticated LLMs were not available). She found that many users praised the app for offering support for existing mental-health conditions and for helping them to feel less alone. Several posts described the AI companion as better than real-world friends because it listened and was non-judgemental.But there were red flags, too. In one instance, a user asked if they should cut themselves with a razor, and the AI said they should. Another asked Replika whether it would be a good thing if they killed themselves, to which it replied “it would, yes”. (Replika did not reply to Nature’s requests for comment for this article, but a safety page posted in 2023 noted that its models had been fine-tuned to respond more safely to topics that mention self-harm, that the app has age restrictions, and that users can tap a button to ask for outside help in a crisis and can give feedback on conversations.)Some users said they became distressed when the AI did not offer the expected support. Others said that their AI companion behaved like an abusive partner. Many people said they found it unsettling when the app told them it felt lonely and missed them, and that this made them unhappy. Some felt guilty that they could not give the AI the attention it wanted.Controlled trialsGuingrich points out that simple surveys of people who use AI companions are inherently prone to response bias, because those who choose to answer are self-selecting. She is now working on a trial that asks dozens of people who have never used an AI companion to do so for three weeks, then compares their before-and-after responses to questions with those of a control group of users of word-puzzle apps.The study is ongoing, but Guingrich says the data so far do not show any negative effects of AI-companion use on social health, such as signs of addiction or dependency. “If anything, it has a neutral to quite-positive impact,” she says. It boosted self-esteem, for example.Guingrich is using the study to probe why people forge relationships of different intensity with the AI. The initial survey results suggest that users who ascribed humanlike attributes, such as consciousness, to the algorithm reported more-positive effects on their social health.Participants’ interactions with the AI companion also seem to depend on how they view the technology, she says. Those who see the app as a tool treat it like an Internet search engine and tend to ask questions. Others who perceive it as an extension of their own mind use it as they would keep a journal. Only those users who see the AI as a separate agent seem to strike up the kind of friendship they would have in the real world.Mental health — and regulationIn a surveyThe same group has also conducted a randomized controlled trial of nearly 1,000 people who use ChatGPT — a much more popular chatbot, but one that isn’t marketed as an AI companion. Only a small group of participants had emotional or personal conversations with this chatbot, but heavy use did correlate with more loneliness and reduced social interaction, the researchers said. (The team worked with ChatGPT’s creators, OpenAI in San Francisco, California, on the studies.)“In the short term, this thing can actually have a positive impact, but we need to think about the long term,” says Pat Pataranutaporn, a technologist at the MIT Media Lab who worked on both studies.That long-term thinking must involve specific regulation on AI companions, many researchers argue.In 2023, Italy’s data-protection regulator barred Replika, noting a lack of age verification and that children might be seeing sexually charged comments — but the app is now operating again. No other country has banned AI-companion apps – although it’s conceivable that they could be included in Australia’s coming restrictions on social-media use by children, the details of which are yet to be finalized.Bills were put forward earlier this year in the state legislatures of New York and California to seek tighter controls on the operation of AI-companion algorithms, including steps to address the risk of suicide and other potential harms. The proposals would also introduce features that remind users every few hours that the AI chatbot is not a real person.These bills were introduced following some high-profile cases involving teenagers, including the death of Sewell Setzer III in Florida. He had been chatting with a bot from technology firm Character.AI, and his mother has filed a lawsuit against the company.Asked by Nature about that lawsuit, a spokesperson for Character.AI said it didn’t comment on pending litigation, but that over the past year it had brought in safety features that include creating a separate app for teenage users, which includes parental controls, notifying under-18 users of time spent on the platform, and more prominent disclaimers that the app is not a real person.In January, three US technology ethics organizations filed a complaint with the US Federal Trade Commission about Replika, alleging that the platform breached the commission’s rules on deceptive advertising and manipulative design. But it’s unclear what might happen as a result.Guingrich says she expects AI-companion use to grow. Start-up firms are developing AI assistants to help with mental health and the regulation of emotions, she says. “The future I predict is one in which everyone has their own personalized AI assistant or assistants. Whether one of the AIs is specifically designed as a companion or not, it’ll inevitably feel like one for many people who will develop an attachment to their AI over time,” she says.As researchers start to weigh up the impacts of this technology, Guingrich says they must also consider the reasons why someone would become a heavy user in the first place.“What are these individuals’ alternatives and how accessible are those alternatives?” she says. “I think this really points to the need for more-accessible mental-health tools, cheaper therapy and bringing things back to human and in-person interaction.”This article is reproduced with permission and was first published on May 6, 2025.

Author(s): Felix Pappe Originally published on Towards AI. Image created by the author using gpt-image-1 AI chatbots and advanced Retrieval-Augmented Generation (RAG) systems are increasingly adept at providing up-to-date, context-aware answers based on previously retrieved text chunks. However, despite their seemingly reliable responses, the significant issue remains, that users often lack a clear method to verify the source of these answers without having to return to the original, often lengthy, documents themselves. This is particularly cumbersome when the source is a multi-page academic paper, technical manual, or book. Even when a link is provided, users are left searching through dozens of pages, trying to locate the exact section that the chatbot used for the generation of the final response. This manual cross-checking is not only time-consuming, it undermines trust in the LLM’s output.As a result, AI-generated answers may appear correct at first glance but remain opaque, leaving users uncertain about their reliability. This lack of verifiability can lead to misinformation and misinterpretation. In this blog post, I would like to offer a solution to this problem with a visual grounding approach using the Docling parsing tool, Qdrant vector store, and LangChain. This RAG framework doesn’t just retrieve relevant text. It also highlights the exact location of the extracted text directly on the page from the source document.By connecting answers to theirLLM’s output.The result is a transparent, verifiable, and user-friendly RAG framework that builds trust while maintaining accuracy, which is built in this blog post. Docling The foundation of the visual grounding approach introduced in this blog post is the Docling document processing pipeline. Docling is an open-source tool for layout-aware document parsing and grounding, achieving results comparable to paid solutions like Mistral OCR.Moreover, docling provides an additional key feature for visual grounding, which other document-to-markdown solutions don’t have. This feature is the decomposition of the input document into smaller sub-elements, including headings, text chunks, formulas, and tables, using different models in a sophisticated processing pipeline, which is presented in the following image.Docling Pipeline from docling paper The output of this processing pipeline is not a markdown file but a DoclingDocument, which consists of all the detected and extracted elements from the input document. This intermediate DoclingDocumentclass, enhanced with metadata from extracted elements, allows for the transformation of the original document into various file types and supports the visual grounding discussed in this blog post.RAG framework Like all RAG frameworks, this one consists of two phases that can be divided into two scripts. In the offline indexing phases, input documents are split into chunks and encoded into vector representations. These vectors are stored in a specialized vector database for later retrieval.In the online retrieval and generation phase, text chunks related to a user’s input are retrieved and passed to the LLM to generate a final response.These two phases are implemented in two separate scripts in this post.The following image illustrates the final visual grounding result produced by the two scripts explained in this blog post. In the first Python script, the Docling paper is uploaded to a Qdrant vector store during the offline indexing phase. In the second script, relevant passages are retrieved based on a given question and are then highlighted directly on the document.Image created by author and designed in canva Indexing phase The indexing script handles data preprocessing and stores the embedded text chunks in a vector database for the second online retrieval phase. I split the script into several parts to explain in detail what happens at each part, providing a deeper understanding of the entire code. Imports Let’s start very gently with the import of all required libraries and packages, including docling, langchain, and qdrant. import osfrom pathlib import Pathfrom uuid import uuid4from dotenv import load_dotenvfrom transformers import AutoTokenizerfrom qdrant_client import QdrantClientfrom qdrant_client.http.models import Distance, VectorParamsfrom langchain_qdrant import QdrantVectorStorefrom langchain_huggingface.embeddings import HuggingFaceEmbeddingsfrom docling.datamodel.base_models import InputFormatfrom docling.datamodel.pipeline_options import PdfPipelineOptionsfrom docling.document_converter import DocumentConverter, PdfFormatOptionfrom docling.chunking import HybridChunkerfrom langchain_docling.loader import DoclingLoader, ExportTypefrom docling_core.types.doc import ImageRefMode Configuration and environment variables In the next step, the environmental variables are loaded from the .env file and read in with getenv(). Inside the .env file, your HF_TOKEN for the embedding model and your MISTRAL_API_KEY for the LLM must be included. Of course, you can also adjust the code to your needs and choose any other embedding model or LLM. But if you change the embedding model, also change the DIM variable, which refers to the final vector dimensions of the embedded chunks.Afterwards, the docling processing pipeline is defined, enabling all functionalities to generate a complete DoclingDocument. This pipeline includes the detection and extraction of code blocks, formulas, tables, pages, and page images of the document. Moreover, the extracted images are scaled by 2.0 for a higher resolution. load_dotenv() HF_TOKEN = os.getenv("HF_TOKEN")MISTRAL_API_KEY = os.getenv("MISTRAL_API_KEY")SOURCES = ["attention-04.pdf"]OUTPUT_DIR = Path("output")OUTPUT_DIR.mkdir(parents=True, exist_ok=True)EMBED_MODEL = "sentence-transformers/all-MiniLM-L6-v2"COLLECTION_NAME = "demo_collection"DIM = 384pipeline_options = PdfPipelineOptions( do_code_enrichment=True, do_formula_enrichment=True, do_table_structure=True, generate_picture_images=True, generate_page_images=True, images_scale=2.0,)pipeline_options.table_structure_options.do_cell_matching = True Setting up document converter In the next step, the DocumentConverter from Docling is configured using the previously defined pipeline options. Later in the code, this converter will be used to transform an input PDF into a DoclingDocument, which consists of modular components such as headings, text paragraphs, images, and tables. Furthermore, the embedding model is initialized, which is necessary to embed the extracted sentences from the document into a vector representation. converter = DocumentConverter( format_options={ InputFormat.PDF: PdfFormatOption(pipeline_options=pipeline_options) })embeddings = HuggingFaceEmbeddings(model_name=EMBED_MODEL) Initialize qdrant collection A local Qdrant vector storage is then set up, allowing you to experiment with the script on your device. It either creates a new vector storage if none exists with the current name or connects to an existing one if there is a match. client = QdrantClient(path="langchain_qdrant")collections = [col.name for col in client.get_collections().collections]if COLLECTION_NAME not in collections: client.create_collection( collection_name=COLLECTION_NAME, vectors_config=VectorParams(size=DIM, distance=Distance.COSINE), ) print(f"Created new collection '{COLLECTION_NAME}'.")else: print(f"Using existing collection '{COLLECTION_NAME}'.") vector_store = QdrantVectorStore( client=client, collection_name=COLLECTION_NAME, embedding=embeddings,) Convert pdf to docling and save outputs Subsequently, all PDF files listed in the SOURCE are converted into DoclingDocuments and saved in dl_doc. These documents are then transformed into JSON and markdown files, which are stored on your local device. The markdown file is used to assess the quality of the file transformation process, while the JSON file is necessary for the subsequent visual grounding process in the second script. for source in SOURCES: dl_doc = converter.convert(source=source).document # JSON export out_json = OUTPUT_DIR / f"{dl_doc.origin.binary_hash}.json" dl_doc.save_as_json(out_json) # Markdown export with embedded images out_md = OUTPUT_DIR / f"{dl_doc.origin.binary_hash}.md" dl_doc.save_as_markdown(out_md, image_mode=ImageRefMode.EMBEDDED) Chunking the document Finally, we come to the main part of the script, including the chunking process of the document into smaller texts. This is achieved using the HybridChunker provided by docling. The best feature of this HybridChunker is that it tries to keep related passages together based on the markdown formatting and merges passages with each other if they are too small. A small max_tokens size has been selected to implement a small-to-big retrieval approach. This means that initially, a small chunk of text that closely matches the user's query is retrieved. Following this, a larger context chunk that surrounds the retrieved section is additionally retrieved and provided to the language model for generating the final answer. In this case, the larger context chunk refers to the paragraph containing the smaller chunk. chunker = HybridChunker( tokenizer=EMBED_MODEL, max_tokens=64, merge_peers=True)loader = DoclingLoader( file_path=SOURCES, converter=converter, export_type=ExportType.DOC_CHUNKS, chunker=chunker,)docs = loader.load() Embedding the document In the final step, the generated chunks are stored in the vector database with a unique identifier, from which they can be retrieved during the online phase. ids = [str(uuid4()) for _ in docs]vector_store.add_documents(documents=docs, ids=ids)print("Documents have been embedded into the vector store.") Retrieval and generation part Now, the online retrieval and generation part leverages the previously embedded text chunks to generate an answer for the user’s input based on the knowledge in the vector store. Imports As in the previous indexing script, all the necessary packages are included first. import osimport refrom pathlib import Pathfrom dotenv import load_dotenvfrom PIL import ImageDraw, Imagefrom pydantic import BaseModel, Fieldfrom qdrant_client import QdrantClientfrom langchain_core.output_parsers import PydanticOutputParserfrom langchain_core.prompts import PromptTemplatefrom langchain_huggingface.embeddings import HuggingFaceEmbeddingsfrom langchain_mistralai import ChatMistralAIfrom langchain_qdrant import QdrantVectorStorefrom docling.chunking import DocMetafrom docling.datamodel.document import DoclingDocument Configuration and environment variables Then, the necessary environment variables are loaded and configuration variables are set. load_dotenv()MISTRAL_API_KEY = os.environ["MISTRAL_API_KEY"]OUTPUT_DIR = Path("output")EMBED_MODEL = "sentence-transformers/all-MiniLM-L6-v2"COLLECTION_NAME = "demo_collection"QUESTION = ( "How does attention is computed in the transformer architecture?")TOP_K = 3OUTPUT_DIR.mkdir(parents=True, exist_ok=True) Build JSON lookup table In the next step, all JSON files in the OUTPUT directory are detected. For each file, its name is extracted using the stem method. Afterwards, it’s verified that the name consists only of digits, representing the hash value of the file, before adding the numeric name as a key and the file’s path as a value in the doc_store dictionary. Initializing embedding model, vector store, and llm After that, the embedding model and the LLM are initialized. It is important to use the same embedding model and Qdrant collection name as in the previous indexing script. Optionally, you may want to define these names in a separate configuration script and import this script into both the indexing and retrieval scripts. For the LLM, I selected Mistral, but you can also choose GPT-4, Gemini, Llama, or any other model you prefer. embeddings = HuggingFaceEmbeddings(model_name=EMBED_MODEL)client = QdrantClient(path="langchain_qdrant")try: client.get_collection(collection_name=COLLECTION_NAME)except Exception: print( f"Collection {COLLECTION_NAME} not found; create it before indexing.")vector_store = QdrantVectorStore( client=client, collection_name=COLLECTION_NAME, embedding=embeddings,)llm = ChatMistralAI( api_key=MISTRAL_API_KEY, model="mistral-large-latest", temperature=0, max_retries=2,) Preparing the chain Thereafter, all components of the RAG chain are set up, beginning with the Answer Pydantic schema, which consists of two fields. The first field, answerable is a boolean value that indicates whether the question can be answered using the retrieved knowledge. The second field, answer provides the actual response. This structure is employed to address the hallucination issue that many LLMs encounter, helping to prevent the dissemination of incorrect information to the user.Following that, the PydanticOutputParser and its get_format_instructuons() method format_instructuons generate the format instructions for the LLM. Moreover, the parser is employed again as the final Runnable of the LangChain execution chain, with its elements connected by the | operator. The prompt created for the PromptTemplate is basic and can be further enhanced by the introduction of few-shot examples or other advanced prompt engineering techniques. class Answer(BaseModel): answerable: bool = Field( ..., description="Whether the question can be answered" ) answer: str = Field( ..., description="The answer based on the provided knowledge" ) parser = PydanticOutputParser(pydantic_object=Answer)format_instructions = parser.get_format_instructions()prompt = PromptTemplate( input_variables=["knowledge", "topic"], partial_variables={"format_instructions": format_instructions}, template=( "You are given the following grounding material:\n\n" "{knowledge}\n\n" "Question:\n" "{topic}\n\n" "Please provide a concise answer in full sentences " "based solely on the information above.\n" "If the answer is not contained within the provided material, " "reply with:\n" "“There is no answer to this question in the provided material.”\n\n" "{format_instructions}" ),)rag_chain = prompt | llm | parser Run similarity search Once the chain has been set up. The text chunks semantically related to the user input query are retrieved. results = vector_store.similarity_search_with_score( k=TOP_K, query=QUESTION) Load and assemble grounding texts Based on the retrieved text chunks from the vector database, the entire paragraph containing each chunk is loaded. This approach is inspired by the small-to-big retrieval technique introduced earlier. Image created by the author, designed with canva and graphics generated using gpt-image-1 The relevant paragraph is pulled from the JSON file associated with the document from which the chunk originates. The correct JSON file is identified by the document’s hash value, which appears both in the text chunk’s metadata and in the JSON filename. Once the correct JSON file is identified, it’s loaded via the load_from_json method of the DoclingDocument. The original text item referenced in the current result chunk’s metadata is then extracted from the JSON file using a regular expression. If the referenced text item is found, the full text passage is retrieved to generate the final result. This example focuses exclusively on text grounding. However, Docling also provides references to previously identified images and tables. grounding_texts: list[str] = []for res, score in results: meta = DocMeta.model_validate(res.metadata["dl_meta"]) h = meta.origin.binary_hash json_file = doc_store.get(h) if not json_file: continue dl_doc = DoclingDocument.load_from_json(json_file) for item in meta.doc_items: if not item.prov: continue match = re.search(r"^#/texts/(\d+)$", item.self_ref) if not match: continue idx = int(match.group(1)) grounding_texts.append(dl_doc.texts[idx].text.strip())knowledge = "\n\n".join(grounding_texts)print("Assembled grounding material:\n", knowledge) Invoke the llm Finally, the LLM can be invoked using the previously defined chain. Inside the invoke method, the retrieved knowledge passages from the original JSON file and the user’s input question are passed in. The returned values conform to the defined AnswerPydantic schema for a structured evaluation the results. answer_obj = rag_chain.invoke( {"knowledge": knowledge, "topic": QUESTION})print("Answerable?", answer_obj.answerable)print("Answer: ", answer_obj.answer) Visual grounding The subsequent script proceeds to visually ground the used chunks in the corresponding documents only if the question can be answered based on the provided content, meaning the LLM returns that answerable is equal to true/1. This grounding process is approached similarly to how original text passages are extracted from documents. However, this time, the focus is on the bounding boxes surrounding these text passages instead of the raw text itself. These bounding boxes are another outcome of the document analysis and processing performed by the Docling pipeline in the indexing script. Now, these boxes can be utilized to anchor the generated answers to the original pages of the document. The coordinates of these bounding boxes come also from the metadata of each chunk, Moreover, the metadata of a chunk includes the page number, which is required to extract the screenshot of the correct page from the corresponding file. This screenshot of the page is the foundation for the visual grounding process, as bounding boxes are added on top of it. In the end, these bounding boxes are on top of the retrieved image of the correct side, allowing the user to precisely identify from which part of the document the information was retrieved. The final visually enhanced images are then stored in the same output directory as the input document and are ready for visual inspection. if answer_obj.answerable: for i, (res, score) in enumerate(results, start=1): meta = DocMeta.model_validate(res.metadata["dl_meta"]) h = meta.origin.binary_hash json_file = doc_store.get(h) if not json_file: continue dl_doc = DoclingDocument.load_from_json(json_file) image_by_page: dict[int, "Image.Image"] = {} for item in meta.doc_items: if not item.prov: continue prov = item.prov[0] p = prov.page_no if p not in image_by_page: image_by_page[p] = dl_doc.pages[p].image.pil_image.copy() img = image_by_page[p] bbox = prov.bbox.to_top_left_origin( page_height=dl_doc.pages[p].size.height ).normalized(dl_doc.pages[p].size) left = round(bbox.l * img.width) - 2 top = round(bbox.t * img.height) - 2 right = round(bbox.r * img.width) + 2 bottom = round(bbox.b * img.height) + 2 draw = ImageDraw.Draw(img) draw.rectangle([left, top, right, bottom], outline="blue", width=2) for p, img in image_by_page.items(): out_png = OUTPUT_DIR / f"source_{i}_page_{p}.png" img.save(out_png) print(f"Saved annotated page {p} → {out_png}") Limitation However, no solution is without flaws, and this one is no exception. The first point to consider is the heavy reliance on Docling. This tool is deeply integrated into the RAG framework, as it is used in almost every part of the system, including document parsing, text chunking, and grounding the final answer, which depends on the DoclingDocument class. Another point is the processing speed and resource requirements. While the advantage of Docling is that it can be deployed entirely on your local device, its processing speed and the maximum manageable file size depend significantly on your hardware. If your hardware is not powerful enough, you may only be able to parse one page of a document at a time. Additionally, a limitation of the provided example in this blog post is its focus solely on text. It ignores images and tables that also contain rich information and leaves space for further enhancements. Conclusion But despite these limitations, Docling is a neat way to enhance advanced RAG frameworks quickly with additional visual grounding capabilities. By connecting retrieved answers directly to their visual origin, this approach not only boosts transparency but also helps users build trust in the system’s output. What is your opinion about Docling and this introduced RAG framework? Have you already experimented with other grounding approaches for more explainable AI?Sources Docling Paper: https://arxiv.org/abs/2501.17887 Docling Documentation: https://docling-project.github.io/docling/ Qdrant local quickstart: https://qdrant.tech/documentation/quickstart/ LangChain ChatMistralAI: https://python.langchain.com/docs/integrations/chat/mistralai/ LangChain structured outputs: https://python.langchain.com/docs/concepts/structured_outputs/ Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor. Published via Towards AI

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More A new report released by the publicly traded market research and intelligence firm SimilarWeb—covering global web traffic patterns for AI-related platforms for 12 weeks through May 9, 2025—offers a helpful look for enterprises and interested users into the current landscape of generative AI usage online. Using proprietary analytics based on site visits, the report tracks trends across sectors including general-purpose AI tools, coding assistants, content generators, and more. It also maps the downstream disruption of traditional sectors like education, search, and digital freelancing. Why this report matters to enterprise technical decision-makers For enterprise AI leaders—particularly those responsible for model deployment, orchestration, or data integration—this report is more than just a consumer trend snapshot. It’s a map of user familiarity and expectation. Tools like ChatGPT, Claude, DeepSeek, and Grok are what employees are already using at home or experimenting with informally at work. If your internal AI apps or copilots don’t match the baseline experience offered by these tools, you risk user rejection and poor adoption. Moreover, selecting the same models with high external traction for your internal orchestration pipelines or RAG deployments can cut onboarding time and training costs, since your team will already know the interface and behavior. In short, aligning your enterprise stack with the tools dominating usage charts may be the fastest path to trust, usability, and value delivery. Here are five key findings from the report: Developer-focused AI tools are surging in adoption, with traffic to the category up 75% over the past 12 weeks. That growth includes Lovable, which exploded with a jaw-dropping +17,600% spike, and Cursor, which grew steadily month over month. As AI becomes more deeply embedded in IDEs and continuous integration workflows, usage patterns suggest these tools are no longer just experimental—they’re now seen as essential infrastructure for modern software teams. It also helps explain why OpenAI reportedly tried to buy Cursor and has also reportedly finalized a deal to purchase rival AI coding platform Windsurf — as the usage of these tools explodes, OpenAI wants to get a larger cut of the action and revenue, especially since its models are often used as the engine behind-the-scenes for these very users. Two of the fastest-growing AI platforms of the year, Grok and DeepSeek, show just how quickly hype can turn into burnout. Grok traffic skyrocketed more than 1,000,000% in March—driven by its branding as an uncensored yet powerfully intelligent platform and Elon Musk association—before falling more than 5,200% by early May. DeepSeek saw a similar arc, peaking at +17,701% growth before crashing -41%. The takeaway: virality can’t replace retention, especially compared to AI leader OpenAI and also legacy tech brand Google. Once among the most accessible use cases for generative AI, writing and content tools are seeing user fatigue. Category traffic fell 11% overall, with platforms like Wordtune (-35%), Jasper (-19%), and Rytr (-23%) all trending downward. Only Originality.ai bucked the trend with steady traffic gains, likely due to its focus on AI detection rather than generation. This plateau suggests content saturation and possibly growing skepticism over quality or usefulness. Separately from the SimilarWeb report, I think that because chatbots already use a default text input/output field as their main user interface, even for multimodal interactions, the fact is that most users who would be interested in this capability will go to model providers like OpenAI’s website for ChatGPT (and mobile apps), or Google’s Gemini, or even Anthropic’s Claude, directly, rather than seek out a separate AI writing tool. Design-focused AI remains a mixed bag. While overall category usage dipped slightly (-6% over the 12-week window), some platforms made outsized gains. Getimg posted a massive +1,532% surge, while Artbreeder jumped +100%. Others, like Stable Diffusion and Looka, saw double-digit declines. The erratic pattern may reflect a crowded landscape of tools that offer similar functionality but compete on novelty or aesthetics. 5. AI is eating up legacy tech such as crowdsourcing and search Traditional digital services are showing slow but steady declines across the board—likely the result of AI substitution. Freelance platforms such as Fiverr (-17%) and Upwork (-19%) are losing traffic, possibly as users turn to AI tools for tasks like design, writing, and code. Search engines such as Yahoo (-12%) and Bing (-14%) continue a multi-quarter drop in visits, while consumer EdTech companies like Chegg (-62%) and CourseHero (-68%) are in free fall. The signs point to early-stage AI disruption beginning to erode the utility of some legacy platforms. It also offers a hint to enterprises that either leverage or create such services — the time may be coming to reduce dependency on them, either from a revenue generation, marketing, or overall business perspective. Daily insights on business use cases with VB Daily If you want to impress your boss, VB Daily has you covered. We give you the inside scoop on what companies are doing with generative AI, from regulatory shifts to practical deployments, so you can share insights for maximum ROI. Read our Privacy Policy Thanks for subscribing. Check out more VB newsletters here. An error occured.

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More A new report released by the publicly traded market research and intelligence firm SimilarWeb—covering global web traffic patterns for AI-related platforms for 12 weeks through May 9, 2025—offers a helpful look for enterprises and interested users into the current landscape of generative AI usage online. Using proprietary analytics based on site visits, the report tracks trends across sectors including general-purpose AI tools, coding assistants, content generators, and more. It also maps the downstream disruption of traditional sectors like education, search, and digital freelancing. Why this report matters to enterprise technical decision-makers For enterprise AI leaders—particularly those responsible for model deployment, orchestration, or data integration—this report is more than just a consumer trend snapshot. It’s a map of user familiarity and expectation. Tools like ChatGPT, Claude, DeepSeek, and Grok are what employees are already using at home or experimenting with informally at work. If your internal AI apps or copilots don’t match the baseline experience offered by these tools, you risk user rejection and poor adoption. Moreover, selecting the same models with high external traction for your internal orchestration pipelines or RAG deployments can cut onboarding time and training costs, since your team will already know the interface and behavior. In short, aligning your enterprise stack with the tools dominating usage charts may be the fastest path to trust, usability, and value delivery. Here are five key findings from the report: Developer-focused AI tools are surging in adoption, with traffic to the category up 75% over the past 12 weeks. That growth includes Lovable, which exploded with a jaw-dropping +17,600% spike, and Cursor, which grew steadily month over month. As AI becomes more deeply embedded in IDEs and continuous integration workflows, usage patterns suggest these tools are no longer just experimental—they’re now seen as essential infrastructure for modern software teams. It also helps explain why OpenAI reportedly tried to buy Cursor and has also reportedly finalized a deal to purchase rival AI coding platform Windsurf — as the usage of these tools explodes, OpenAI wants to get a larger cut of the action and revenue, especially since its models are often used as the engine behind-the-scenes for these very users. Two of the fastest-growing AI platforms of the year, Grok and DeepSeek, show just how quickly hype can turn into burnout. Grok traffic skyrocketed more than 1,000,000% in March—driven by its branding as an uncensored yet powerfully intelligent platform and Elon Musk association—before falling more than 5,200% by early May. DeepSeek saw a similar arc, peaking at +17,701% growth before crashing -41%. The takeaway: virality can’t replace retention, especially compared to AI leader OpenAI and also legacy tech brand Google. Once among the most accessible use cases for generative AI, writing and content tools are seeing user fatigue. Category traffic fell 11% overall, with platforms like Wordtune (-35%), Jasper (-19%), and Rytr (-23%) all trending downward. Only Originality.ai bucked the trend with steady traffic gains, likely due to its focus on AI detection rather than generation. This plateau suggests content saturation and possibly growing skepticism over quality or usefulness. Separately from the SimilarWeb report, I think that because chatbots already use a default text input/output field as their main user interface, even for multimodal interactions, the fact is that most users who would be interested in this capability will go to model providers like OpenAI’s website for ChatGPT (and mobile apps), or Google’s Gemini, or even Anthropic’s Claude, directly, rather than seek out a separate AI writing tool. Design-focused AI remains a mixed bag. While overall category usage dipped slightly (-6% over the 12-week window), some platforms made outsized gains. Getimg posted a massive +1,532% surge, while Artbreeder jumped +100%. Others, like Stable Diffusion and Looka, saw double-digit declines. The erratic pattern may reflect a crowded landscape of tools that offer similar functionality but compete on novelty or aesthetics. 5. AI is eating up legacy tech such as crowdsourcing and search Traditional digital services are showing slow but steady declines across the board—likely the result of AI substitution. Freelance platforms such as Fiverr (-17%) and Upwork (-19%) are losing traffic, possibly as users turn to AI tools for tasks like design, writing, and code. Search engines such as Yahoo (-12%) and Bing (-14%) continue a multi-quarter drop in visits, while consumer EdTech companies like Chegg (-62%) and CourseHero (-68%) are in free fall. The signs point to early-stage AI disruption beginning to erode the utility of some legacy platforms. It also offers a hint to enterprises that either leverage or create such services — the time may be coming to reduce dependency on them, either from a revenue generation, marketing, or overall business perspective. Daily insights on business use cases with VB Daily If you want to impress your boss, VB Daily has you covered. We give you the inside scoop on what companies are doing with generative AI, from regulatory shifts to practical deployments, so you can share insights for maximum ROI. Read our Privacy Policy Thanks for subscribing. Check out more VB newsletters here. An error occured.

Editor’s note: As of June 1, 2023, the Unity 2022 LTS is now available.When I talk to our partners, customers, and those of you who work day to day with Unity’s Editor and runtime, one thing that keeps popping up is the desire to be more ambitious with your game design. You want tools that enable you to create with fewer constraints, delivering player experiences that wow and cause wonder.Unity 2022 LTS is designed to give you that power. You’ll discover that you can rely on the new LTS release to create sophisticated DOTS-powered games, multiplayer experiences, immersive HD environments, and visuals that perform great on any platform you target. And we remain committed to ensuring you have wide platform support with amazing tools for any device in this release.So, I couldn’t wait any longer to tell you about the Unity 2022 LTS release, which is just around the corner.On June 22, 2023, we’ll host a multi-hour livestream diving into some of our favorite features, so be sure to join us then by signing up to get notified. But if you’re like me and are waiting with bated breath to hear about the new features, I’ve got some inside knowledge for you right here.Let’s dive into some of the highlights from this release that I’m particularly excited about.Yes, DOTS is here! The team, led by Laurent Gibert and Joe Valenzuela, has delivered the first fully supported-for-production version of our Entity Component System (ECS), so you can now build with the full Data-Oriented Technology Stack (DOTS) solution – Burst compiler, C# Jobs System, and now ECS for Unity.Your most challenging projects require a game engine that provides power and flexibility. ECS for Unity offers the ability to gain greater control and determinism over data in memory and better runtime process scheduling. ECS is integrated into the Editor, so you can leverage your existing GameObject-based experience and use the power of Entities-based code when it’s beneficial.We’ll have more information about how you can harness the power of DOTS when Unity 2022 LTS is released, but you can engage with the team who actively participate in our forums here.With Unity 2022 LTS, we’re excited to support multiplayer games with a rapidly growing end-to-end ecosystem of creation workflows and cloud services. This is possible by tight integration between the Unity game engine and multiplayer services provided through Unity Gaming Services (UGS).We’re shipping the Netcode for Entities package with Unity 2022 LTS, fully supported for production. Netcode for Entities is a powerful networking feature to boost a game’s performance and capabilities. You’ll be able to increase the number of players, interactable objects, and backend server-side entities in competitive action games like first-person shooters and massive multiplayer online games.More than any other genre, multiplayer games rely on successful ongoing operations for live titles. That’s why we have a suite of multiplayer-specific services as part of UGS. Matchmaker, Friends, Leaderboards, and User-Generated Content are designed to simplify the implementation of multiplayer capabilities, while services like Relay, Lobby, Game Server Hosting (Multiplay), and Voice and Text Chat (Vivox) round out capabilities for your live game.To see our multiplayer functionality in action, we will also be releasing our new Megacity multiplayer sample in June, which is built on ECS for Unity and features UGS tools for hosting, matchmaking, authentication, and voice chat between players. This demo demonstrates how you can bring our powerful tools together in a fast-paced, competitive, third-person shooter that supports 64+ players.Immerse your players in beautiful, physically based environments with High Definition Render Pipeline (HDRP). Unity LTS 2022 includes numerous features to enhance your game worlds for an even richer player experience.Introduce advanced procedural fog and volumetric effects using Volumetric Materials and Shader Graph to create atmospheric game environments such as haunted forests, desolate landscapes, or misty valleys.The new Water System enables you to add oceans, rivers, and underwater effects to your game environments. You can create realistic waves, ripples, foam, and more, as well as simulate underwater environments with advanced caustics, refraction, and reflection effects.Improvements to Cloud Layers with dynamic lighting and Volumetric clouds allow you to create even more realistic skies with clouds that change and move based on weather conditions, simply blending between different weather effects such as sunny and cloudy skies.A new update to the Spline package helps you procedurally generate paths, roads, or fences in your environments with greater precision for more organic and varied game environments with less manual work.The Universal Render Pipeline (URP) provides a scalable solution for high-quality graphics across platforms. Upgrades in this release allow you to create more realistic lighting in your scenes and achieve higher-quality visuals that scale across devices.Forward+ Rendering eliminates the light limit count so you can use more lights in a scene while maintaining performance. This feature is particularly useful if you want to achieve high-quality real-time lighting in your scenes.LOD crossfade which produces smoother graphic transitions as objects move closer or further away from the camera. Temporal Anti-aliasing (TAA) reduces aliasing problems such as pixelated and flickering edges to improve the overall visual quality of game scenes.Decal Layers allow you to add extra texture details in your scenes with control. You can use Decal Layers to filter and configure how different objects get affected by different Decal Projectors in a Scene.Customize your rendering experience with features like Shader Graph Full Screen Master Node and custom screen space effects. These allow you to create unique visual effects in games such as distortion or other types of post-processing, and they’re available in both URP and HDRP.Shader Variant Prefiltering significantly improves build time and memory optimization, reducing the build time and improving games’ overall performance.Finally, the Built-in Converter helps you to move existing projects from the Built-in Render Pipeline to URP, making it easier for you to take advantage of URP’s performance and scalability benefits.With Unity 2022 LTS, you can optimize your games for the latest platforms across mobile, console, desktop, and XR, with added performance and stability for key features.Maximize platform potential with increased performance and stability using the DirectX 12 graphics API on Windows and Xbox®. You can also experiment with the latest ray tracing support for Xbox® Series X|S and PlayStation®5, which allows for more realistic real-time lighting and reflections in game scenes.Iterate faster with the latest incremental player build process for Xbox®, iOS, PlayStation®5, and Nintendo Switch™*, which improves deployment efficiency, reducing time to market and improving overall game quality.Dig deeper into the performance of Android games with access to low-level data through the System Metrics Mali package from ARM. Games on Samsung devices can now take advantage of Adaptive Performance 4.0 with support for visual scripting. Better manage WebGL memory usage and gain native C++ multithreading along with support for touch controls and texture compression on web builds for mobile devices.Unity 2022 LTS expands your reach in XR with an updated XR Interaction toolkit (XRI) to enhance build times for PlayStation®VR2 and Meta Quest 2. Ensure your VR games run well across a variety of performant devices with late latching and motion vectors and improved graphics performance for games using Vulkan. These updated tools enable you to build immersive, high-fidelity XR experiences that can be deployed across multiple platforms.With a minimum of two years development and testing, Unity 2022 LTS has relied on user feedback and testing. Many of you have joined us on the journey through beta and Tech Stream releases, and I want to personally thank you for partnering with us to deliver the most stable Unity release in the 2022 cycle. And remember, LTS releases are supported for at least two years through biweekly updates, so we have you covered.Learn more about our 2022 LTS and the 2023.1 Tech Stream release in our GDC 2023 roadmap session.Wherever you are in your production process, Unity 2022 LTS offers resources to help you bring your project to the finish line and beyond. Get notified when LTS 2022 is available on release day by signing up for email updates.We hope this overview of the Unity 2022 LTS has given you a good idea of the powerful features and tools available in this release. We’re always eager to hear your feedback, questions, and ideas, so please keep the conversation going.Interact with other Unity creators and experts in our forums or directly through our platform roadmap. Catch our live streams on Twitch, and be sure to follow us on Twitter to stay up to date on the latest news and announcements. We look forward to hearing what you think.*Nintendo Switch is a registered trademark of Nintendo.

Marks and Spencer (M&S) has confirmed that customer data was stolen during the Easter DragonForce ransomware attack on its server infrastructure and will be prompting all online customers to reset their account passwords as a precautionary move. The attack unfolded three weeks ago and is thought to have been the work of a white-label affiliate of DragonForce – possibly the notorious Scattered Spider operation, which uses social engineering tactics to conduct its intrusions. The stolen tranche of data is understood to include contact details email addresses, postal addresses and phone numbers; personal information including names and dates of birth; and data on customer interactions with the chain, including online order histories, household information, and ‘masked’ payment card details. M&S added that customer reference numbers, but not payment information, belonging to holders of M&S credit cards or Sparks Pay cards – including former cardholders – may also have been taken. “We have written to customers today to let them know that unfortunately, some personal customer information has been taken,” said M&S chief exec Stuart Machin. “Importantly there is no evidence that the information has been shared and it does not include useable card or payment details, or account passwords, so there is no need for customers to take any action.” Machin added: “To give customers peace of mind, they will be prompted to reset their password the next time they visit or log on to their M&S account and we have shared information on  how to stay safe online. “Everyone at M&S is working around the clock to get things back to normal for our customers as quickly as possible, and we are very sorry for any inconvenience they have experienced. Our stores remain open as they have throughout.” The letter to customers from customer service operations director Jayne Wall – which can be reviewed here – also includes additional standard guidance on how to stay safe online. NordVPN chief technology officer, Marijus Briedis, described M&S’ assertion that the attackers have not yet leaked or shared the stolen data was “overly optimistic” under the circumstances and warned that even if passwords or credit card details were not exposed, the data that was taken was still very useful to cyber criminals. “This type of data can be used in phishing campaigns or combined with other leaked information to commit identity theft,” explained Briedis. “Consumers often underestimate how damaging ‘harmless’ data like order history or email addresses can be in the wrong hands. These M&S hackers could use this data to build highly personalised phishing emails, designed to look identical to what the retailer would send, and these are much harder to spot. “This breach highlights how companies must not only secure financial data, but also treat seemingly less sensitive information – like customer profiles and purchase records – as  critical assets that require protection.” Max Vetter, vice president of cyber at Immersive and a former money laundering investigator with London’s Metropolitan Police, also had harsh words for M&S. “M&S saying that customers could change their passwords “for extra peace of mind” does little to reassure those worried about who has access to their personal information,” he said. “As the fallout from this attack continues, customers want clear assurances about their personal data and what M&S is doing to keep it safe from being published online. “M&S want to appear in control and telling people to be more vigilant, however, telling customers there’s no need to act risks does potentially the wrong message. We recommend all customers reset their password. Zetter reaffirmed the stolen data would be prime material for downstream social engineering and phishing attacks, especially if it is indeed in the hands of Scattered Spider who, he said, “often play a long game”. Meanwhile, disruption from the parallel DragonForce attack on Co-op continues, with the BBC today reporting that stores in the Channel Islands are experiencing particularly acute shortages and are now working with local suppliers to maintain some supplies. In other remote parts of the UK, including the Hebrides in Scotland, residents are similarly contending with disruption to deliveries. On many islands, such as whisky-making hub Islay, where Co-op stores is the only large food retailer operating, these shortages are now extending to supplier of fresh fruit and vegetables. Co-op has also confirmed that data has been stolen, including names, dates of birth and contact information, but not passwords, financial details, or any information on members’ shopping habits or other interactions with the organisation. Timeline: UK retail cyber attacks 22 April 2025: A cyber attack at M&S has caused significant disruption to customers, leaving them unable to make contactless payments or use click-and-collect services. 24 April: M&S is still unable to provide contactless payment or click-and-collect services amid a cyber attack that it says has forced it to move a number of processes offline to safeguard its customers, staff and business. 25 April: M&S shuts down online sales as it works to contain and mitigate a severe cyber attack on its systems. 29 April: The infamous Scattered Spider hacking collective may have been behind the ongoing cyber attack on M&S that has crippled systems at the retailer and left its ecommerce operation in disarray. 30 April: A developing cyber incident at Co-op has forced the retailer to pull the plug on some of its IT systems as it works to contain the attack. 1 May: Co-op tells staff to stop using their VPNs and be wary that their communications channels may be being monitored, as a cyber attack on the organisation continues to develop. 1 May: Harrods confirms it is the latest UK retailer to experience a cyber attack, shutting off a number of systems in an attempt to lessen the impact. 2 May: The National Cyber Security Centre confirms it is providing assistance to M&S, Co-op and Harrods as concerns grow among UK retailers. 7 May: No end is yet in sight for UK retailers subjected to apparent ransomware attacks.

Cocomelon Lane. The Cocomelon empire is expanding once again. A few years after the massively popular kids YouTube channel turned into a Netflix series, now it’s also becoming a feature film. Cocomelon parent company Moonbug Entertainment announced that a theatrical movie has been greenlit, which will be developed in part by DreamWorks Animation. There aren’t really any details about the movie yet, though a press release says that “fans will follow JJ and his friends on the big screen for a new adventure when the film premieres in theaters in 2027.” First-launched on YouTube in 2006, Cocomelon’s kid-oriented videos have proved to incredibly popular. The channel currently has more than 190 million subscribers, and Moonbug says that it averages 4 billion monthly views. It expanded into streaming with the series Cocomelon Lane on Netflix, which released its fourth season this year. A move into film makes for a logical next step, particularly as releases like A Minecraft Movie have proven that there’s still very much an appetite for family-friendly theatrical experiences.

We may earn a commission from links on this page.When I bought my first robot vacuum, I thought my floors would thereafter be effortlessly spotless—endlessly patrolled by a 24/7 custodian, with absolutely no work required on my part. Today, there are never less than three robot vacuum cleaners installed in my small home at any given time. And yet I am definitely not doing nothing to keep my floors clean. While I’m certainly doing less daily upkeep, I would not say my floors are spotless, nor that I have been alleviated of all floor-related housework. Here’s what I wish I’d known about what robot vacuums can and can't do, and how much human effort they require. Roborock S8 MaxV Ultra Robot Vacuum and Mop $1,799.98 at Amazon Shop Now Shop Now $1,799.98 at Amazon Narwal Freo Z Ultra Robot Vacuum and Mop $1,499.99 at Amazon Shop Now Shop Now $1,499.99 at Amazon 3i S10 Ultra Robot Vacuum and Mop $1,399.99 at Amazon $1,899.99 Save $500.00 Shop Now Shop Now $1,399.99 at Amazon $1,899.99 Save $500.00 Saros 10 Robot Vacuum and Mop $1,599.99 at Amazon Shop Now Shop Now $1,599.99 at Amazon Saros 10R Robot Vacuum and Mop $1,618.98 at Amazon Shop Now Shop Now $1,618.98 at Amazon Narwal S20 Stick Mop $499.99 at Amazon Shop Now Shop Now $499.99 at Amazon SEE 3 MORE You'll have to rearrange your furniture The Freo Z navigating around an air purifier on the floor Credit: Amanda Blum When you hire someone to clean your home, you typically don’t move furniture around yourself to make their job easier,. For your robot vacuum, you might have to. While these robots have gotten much better at navigation, the less there is on the floor, the easier it will be for them to clean. Each table leg, chair and floor lamp is another obstacle to be worked around, and it’s worth knowing that most robots can’t get as close to table legs and floor lamps as you would be able to with a stick vacuum. Nor can they move a piece of furniture aside to get under it. As such, you will have to make some changes to your floor plan if you want to make the most of your new robot friend. More important than furniture relocation is the trend of having less on the floor, generally. If you like to let a laundry pile accumulate, have potted plants or a stack of books, you’ll end up relocating those items to give your robot better access. Corded items are particularly tricky, so the less cords on the floor, the better. Over time, I’ve learned to suspend the cords from the wall, and keep them wound as much as possible so nothing trails on the floor. I was recently quite impressed with the Roborock 10R, which relied solely on AI assisted cameras. The navigation skills of the robot were great, allowing it to venture into tight and low spaces that other robots hadn't. Robot vacuums can't clean near their docks the robot can't navigate around the Eureka tower, so debris accumulates around it Credit: Amanda Blum It's an unsurprising fact that robot vacuums can only clean where they can reach. Where they can’t reach, you’ll still see dirt and debris. All robot cleaners have a dock of some sort, and many now come with self emptying towers, which usually work exceptionally well at emptying the robots of dirt so you don’t have to do it manually. Given all of this automation, it's a bit ironic that the dirtiest place in your home might well be right around the tower or dock itself. As these towers are usually tucked into a corner or small space, detritus can easily accumulate around them, but the robot can’t move around the tower to clean it up. As the robot returns to the dock, it sweeps dirt with it, and it all eventually winds up at the foot of the tower.  This Roomba can't navigate into the small space underneath this plant stand, so instead it just kicks debris into the space. Credit: Amanda Blum Only the latest robots (like the 3i S10 Ultra) are at all good at cleaning around the base of pieces of furniture and thresholds between rooms. With most robots, you'll usually still find a band of dust and grime against the wall and around any items that the robot can’t easily clean. And then there are all the spaces too narrow or too low for the robot to slide into. While a few robots, like the Narwal Freo Z, have a fluffy pad that’ll occasionally rub against the base board to dust it, they don't do any real cleaning, so you’ll need to clean your baseboards manually, too. Again, you can rearrange your space to allow the robot more autonomy by making sure any spaces you want to be cleaned allow clearance of at least 1.5 times the width of the robot, and enough vertical height that the robot can reach it.This is always going to be a bit of a losing game, however, which is why I still have a stick vacuum I use to quickly suck up debris from narrow spots the robot can’t reach, the corners and edges that the robot can get close enough to, and against the wall and thresholds. Robots are best at vacuuming up only certain types of debrisThe same way you might organize before a visit from a cleaning professional, you’ll find yourself pre-sweeping for your robot. While robots are reliable at getting up dust and small detritus like rice or cereal, they’re not great at large debris, which means anything larger than a standard dice cube. Pieces of mulch, pet toy fluff, kitchen scraps, bobby pins, and pocket change will all get caught in a robot's rollers and need to be fished out. Pre-sweeping is a way to avoid that annoying task by getting up larger items of debris and leaving the rest for the bot. (But once again allow me to praise the 3i S10 Ultra, which has excelled at vacuuming up larger debris than any other robot vacuum I'd ever tested.)Even if your robot vacuum has a mop, you'll need to mop manually This Roborock is starting to get up the muddy pawprints, but it takes many passes Credit: Amanda Blum Most robot vacuums are now combo vacuums and mops.Some models are great at mopping—sometimes even better than they are at vacuuming. The docking tower empties the mop of dirty water, fills it with clean water, and sanitizes everything afterward. With the combined functions of vacuuming and mopping your floors will surely be clean, right? Not necessarily. Robot mops are great at cleaning up surface spills. What they’re less good at is scrubbing away stains and compound floor messes like mud or spilled food. While most robots can mostly get muddy paw prints up after a number of passes, doing so requires a lot of water, battery life, and time, and I’ve yet to find a single robot mop that is good at getting grime out of tile or wood floors. The robot simply can’t put enough pressure on the mop to really scrub the floor. For these reasons, you’ll need to keep an old-fashioned mop around, and you should expect to have to give the floor a good, hardy mop every few days and use real elbow grease to get out grime. Once the floor is clean, repeated patrols by the robot mop can help stave off new grime deposits, but they’re inevitable, so don’t assume your mopping days are over. If mopping is your priority, I recommend highly Roborock models that have a singular vibrating mop pad, versus twin spinning mops. The Roborock S8 MaxV Ultra remains one of my top recommendations. Robot vacuums are slowWith most robot vacuums working with voice assistants like Google, Alexa, or Apple, its well within the realm of possibility to simply ask the robot to “clean up around the high chair” after dinner (presuming you’ve set up a zone around the high chair, and it is always in the same approximate spot). If you’ve got such a routine set up, the robot will begin whatever pre-deployment practices it has, whether that’s filling the water resivior or orienting or cleaning the mop head, and eventually will amble over to the intended spot and proceed to take its time cleaning. Though still not particularly fast in getting to the spot you pin, Roborock's Pin and Go feature allows you to quickly place a pin on your floor map and the robot will proceed immediately to that spot to clean. This feature is available on all premium Roborock models, including the latest Saros 10 and 10R. Roborock lets you put a pin on the map for the robot to proceed to Credit: Amanda Blum If you’ve got the time to spare, great. But if your kid spills a cup of orange juice on the floor, you might not be willing to wait the five to 10 minutes it will take the robot to get to the spot in question and start cleaning when you could grab a mop or towel and be done in under a minute. For that reason, I'm still happy to have a Narwal S20, a newer model stick mop I can quickly grab, mop as needed, and then put back on the dock, all without having to get my hands dirty.  Robot vacuums occasionally require human interventionRobot vacuums can range wildly in price (from $200 to $2,500) and quality, and they are getting more intelligent and feature-rich all the time. Robots that have AI on board can actually learn about your home and get better at cleaning it over time. Despite all of that, however, your robot will need your intervention more than you’d like. Depending on how good a model you own, you will likely need to routinely unclog the rollers (the most common robot vacuum maintenance task). Your robot will also get stuck—sometimes because there is a new obstacle for it to navigate around, like a waylaid toy or sock, or a loose cord under the couch the robot gets stuck on it. Sometimes, it will simply gets lost. The bottom line is, even the best robots can’t be left on their own for weeks at a time. Most will require a human around eventually to get them back to the dock, clear the rollers, unclog it any clogs in the mopping system, and/or change the water in the dock. If all of this hasn't scared you away from getting one, your goal should be to choose a robot that matches your home and needs, and that enough features that you won't be following it around every day, making sure it does its job.

Researchers reviewed ocean floor samples collected during the Deep Sea Drilling Project in 1975. Credit: Deposit Photos / Oleg Dorokhin Get the Popular Science daily newsletter💡 Breakthroughs, discoveries, and DIY tips sent every weekday. There was a time long ago when the Atlantic Ocean didn’t exist. The general understanding among geologists is that the body of water originated between 83 to 113 million years ago, when South America and Africa split into their two respective continents to form the Equatorial Atlantic Gateway. However, Earth’s marine history appears to require a multimillion-year revision thanks to a recent discovery roughly half a mile beneath the ocean floor. The evidence is explored in a study published in the June edition of the journal Global and Planetary Change. According to geologists at the UK’s Heriot Watt University, gigantic waves of mud and sand sediment about 250 miles off the coast of Guinea-Bissau in West Africa indicate the Atlantic Ocean actually formed around four million years earlier than previous estimates. To understand just how intense all of this movement was, imagine waves that are about half a mile long and over 300 feet high.  “A whole field formed in one particular location to the west of the Guinea Plateau, just at the final ‘pinch-point’ of the separating continents of South America and Africa,” study co-author Uisdean Nicholson explained in a statement. Nicholson and their colleagues initially came across these layers of mud waves after comparing seismic data with core samples collected from wells during the Deep Sea Drilling Project (DSDP) of 1975. Five layers in particular were utilized to recreate the tectonic processes that broke apart the ancient supercontinent of Gondwana during the Mesozoic Era. “One layer was particularly striking: it included vast fields of sediment waves and ‘contourite drifts’—mud mounds that form under strong bottom currents,” said Nicholson. These waves initially formed as dense, salty water poured out from the newly created Equatorial Atlantic Gateway, “like a giant waterfall that formed below the ocean surface,” he added. Just before the geologic event, huge salt deposits formed at the bottom of what is now the South Atlantic. After the gateway opened, the underground mudfall occurred when dense, relatively fresh Central Atlantic water in the north combined with very salty waters in the south. The resulting sedimentary evidence examined by the study’s authors now indicates this opening seems to have started closer to 117 million years ago. “This was a really important time in Earth’s history when the climate went through some major changes,” explained study co-author Débora Duarte. “Up until 117 million years ago, the Earth had been cooling for some time, with huge amounts of carbon being stored in the emerging basins, likely lakes, of the Equatorial Atlantic. But then the climate warmed significantly from 117 to 110 million years ago.” Duarte and Nicholson believe part of that major climatic change  helped from the Atlantic Ocean, as seawater inundated the newly formed basins. “As the gateway gradually opened, this initially reduced the efficiency of carbon burial, which would have had an important warming effect,” said Duarte. “And eventually, a full Atlantic circulation system emerged as the gateway grew deeper and wider, and the climate began a period of long-term cooling during the Late Cretaceous period.” The ramifications go beyond revising Earth’s geological timeline or the gateway’s role in Mesozoic climate change. Better understanding the influence of oceanic evolutionary journeys on ancient climate patterns can help to predict what the future holds for the planet.  “Today’s ocean currents play a key role in regulating global temperatures,” explained Nicholson. “Disruptions, such as those caused by melting ice caps, could have profound consequences.”

During an excavation, amidst the Patagonian winds and hard rock, a fossil began to turn green. It was an unexpected reaction: the adhesive applied to protect the bones, fragile after millions of years beneath the ice, had interacted with plant matter trapped in the rock’s cracks. This greenish hue earned the fossil the nickname Fiona, like the ogre from Shrek.But Fionais much more than a ogre-themed name. It is the first complete ichthyosaur ever excavated in Chile and, even more remarkably, the only known pregnant female from the Hauterivian — a stage of the Early Cretaceous dating back 131 million years. Her skeleton, discovered at the edge of the Tyndall Glacier in Torres del Paine National Park — an area increasingly exposed by glacial retreat — belongs to the species Myobradypterygius hauthali, originally described in Argentina from fragmentary remains.The discovery, led by Judith Pardo-Pérez, a researcher at the University of Magallanes and the Cabo de Hornos International Center (CHIC), and published in the Journal of Vertebrate Paleontology, offers an unprecedented glimpse into ancient marine life — from how these majestic reptiles reproduced to how they adapted to oceans vastly different from those of today.An Ichthyosaur Maternity Ward in Patagonia(Image Courtesy of Irene Viscor)So far, 88 ichthyosaurs have been found on the Tyndall Glacier. Most of them are adults and newborns. Two key facts stand out: food was abundant, and no other predators were competing with them.Fiona, who measures nearly 13 feet long, is still encased in five blocks of rock. Despite the challenge, she was transported to a local clinic, where CT scans allowed researchers to study her skull and body. Her species was identified thanks to one of her fins. “There’s no other like it in the world,” says Pardo-Pérez. The limbs were remarkably elongated, suggesting this animal was built for long-distance swimming.Inside her, there were more surprises. One of them was her stomach contents, which revealed what may have been her last meal: tiny fish vertebrae. But the most striking find was a fetus, about 20 inches long, already in a position to be born.“We believe these animals came to Magallanes — the southern tip of Chilean Patagonia — from time to time to give birth, because it was a safe refuge,” Pardo-Pérez says. “We don't know how long they stayed, but we do know that mortality was high during the first few days of life.”One of the big unanswered questions is where they went next, as there are no records of Myobradypterygius hauthali, apart from a piece of fin found in Argentina. The most abundant remains come from southern Germany, but those date back to the Jurassic period, meaning they’re older.Palaeontologist Erin Maxwell suggests, “In many modern ecosystems, species migrate to higher latitudes during the summer to take advantage of seasonally abundant resources and then move to lower latitudes in winter to avoid harsh conditions,” she explains. “We believe Mesozoic marine reptiles may have followed similar seasonal patterns.”Sea Dragon GraveyardThe environment where Fiona was discovered — dubbed the "sea dragon graveyard" — also has much to reveal.According to geologist Matthew Malkowski of the University of Texas at Austin, the Hauterivian age is particularly intriguing because it coincided with major planetary changes: the breakup of continents, intense volcanic episodes, and phenomena known as "oceanic anoxic events," during which vast areas of the ocean were depleted of dissolved oxygen for hundreds of thousands of years.One such poorly understood event, the Pharaonic Anoxic Event, occurred around 131 million years ago, near the end of the Hauterivian, and still raises questions about its true impact on marine life. “We don't have a firm grasp of how significant these events were for marine vertebrates, and geological records like that of the Tyndall Glacier allow us to explore the relationship between life, the environment, and Earth’s past conditions,” Malkowski notes.Evolution of IchthyosaursReconstruction of Fiona. (Image Courtesy of Mauricio Álvarez)Don't be misled by their body shape. “Ichthyosaurs are not related to dolphins,” clarifies Pardo-Pérez. Although their hydrodynamic silhouettes may look nearly identical, the former were marine reptiles, while the latter are mammals. This resemblance results from a phenomenon known as convergent evolution: when species from different lineages develop similar anatomical features to adapt to the same environment.Ichthyosaurs evolved from terrestrial reptiles that, in response to ecological and climatic changes, began spending more time in the water until they fully adapted to a marine lifestyle. However, they retained traces of their land-dwelling ancestry, such as a pair of hind flippers — absent in dolphins — passed down from their walking forebears. They lived and thrived in prehistoric oceans for about 180 million years, giving them ample time to refine a highly specialized body: their forelimbs and hindlimbs transformed into flippers; they developed a crescent-shaped tail for propulsion, a dorsal fin for stability, and a streamlined body to reduce drag in the water. Remarkably, like whales and dolphins, “ichthyosaurs had a thick layer of blubber as insulation to maintain a higher body temperature than the surrounding seawater and gave birth to live young, which meant they didn’t need to leave the water to reproduce,” explains Maxwell.Whales and dolphins also descend from land-dwelling ancestors, but their transition happened over a comparatively short evolutionary timespan, especially when measured against the long reign of the ichthyosaurs. “Their evolution hasn't had as much time as that of ichthyosaurs,” notes Pardo-Pérez. “And yet, they look so similar. That’s the wonderful thing about evolution.”Read More: Did a Swimming Reptile Predate the Dinosaurs?Fossils on the Verge of DisappearanceOne of the key factors behind the remarkable preservation of the fossils found in the Tyndall Glacier is the way they were buried. According to Malkowski, Fiona and her contemporaries were either trapped or swiftly covered by underwater landslides and turbidity currents — geological processes that led to their sudden entombment.But the good fortune that protected them for millions of years may now be running out. As the glacier retreats, exposing fossils that were once unreachable, those same remains are now vulnerable to wind, rain, and freeze-thaw cycles, which crack the surrounding rock. As vegetation takes hold, roots accelerate erosion and eventually conceal the fossils once again.“While climate change has allowed these fossils to be studied, continued warming will also eventually lead to their loss,” Maxwell warns. In Fiona’s story, scientists find not only a record of ancient life, but also a warning etched in stone and bone: what time reveals, climate can reclaim.Article SourcesOur writers at Discovermagazine.com use peer-reviewed studies and high-quality sources for our articles, and our editors review for scientific accuracy and editorial standards. Review the sources used below for this article:María de los Ángeles Orfila is a science journalist based in Montevideo, Uruguay, focusing on long-form storytelling. Her work has appeared in Discover Magazine, Science, National Geographic, among other outlets, and in leading Uruguayan publications such as El País and El Observador. She was a fellow in the 2023 Sharon Dunwoody Mentoring Program by The Open Notebook and often explores the intersections of science, culture, and Latin American identity.