--- title: "Announcing Atlas: The Point-in-time Knowledge Graph" description: "Meet Atlas: 35 million corroborated facts for grounding AI, testing portfolio risk, and mapping supply chains with point-in-time company knowledge." url: "https://nosible.com/blog/announcing-atlas-the-point-in-time-knowledge-graph" --- [NOSIBLE World](https://nosible.com/blog/tag/nosible-world) [Information Retrieval](https://nosible.com/blog/tag/information-retrieval) [Web Search](https://nosible.com/blog/tag/web-search) # Announcing Atlas: The Point-in-time Knowledge Graph [Stuart Reid](https://www.linkedin.com/in/stuartgordonreid/) 2026-09-30 20 min read Copy as Markdown ![Compression is intelligence: NOSIBLE Search indexes the web, World produces events, and Atlas maps facts into a knowledge graph.](https://nosible.com/images/2026/09/atlas-nosible-stack-3640x1560.png) ## Know everything, all the time. NOSIBLE is on a mission to build AIs that know everything, all the time. Today we took a huge step towards achieving that goal. Today we released Atlas, a web-scale, point-in-time knowledge graph. Atlas contains 10 million entities: people (4.7 million), organizations (3.5 million), products (1 million), locations (700,000), and geopolitical entities (500,000). It maps 35 million facts about those entities. Every fact has corroborated, point-in-time evidence. In simple terms, Atlas knows about the World: - Ownership: Who acquired, merged with, or spun off from whom. - Supply: Who supplies, imports from, or distributes for whom. - Legal: Who has sued, investigated, or fined whom. - Capital: Who shorted, downgraded, or underwrote whom. - Policy: Who sanctions, tariffs, or blockades whom. - Threats: Who cyberattacks, boycotts, or strikes against whom. More importantly, Atlas knows exactly WHEN these facts became known. Every fact is corroborated by multiple independent sources as of a given date. Customers can therefore reconstruct exactly what was publicly known at each point in time, inspect the provenance, and ground AI-enabled backtests in contemporaneous evidence to minimize foreknowledge and lookahead bias. We deliver the entire Atlas knowledge graph to your own infrastructure: 10 million entities and 35 million facts, ready to use. You can then add your own internal entities and facts to that private copy. This enables organizations to immediately distinguish true proprietary intelligence from open-source intelligence. Teams can combine public and proprietary intelligence while keeping their own data confidential. Table of contents [Let's hop into Coca-Cola](https://nosible.com/blog/announcing-atlas-the-point-in-time-knowledge-graph#lets-hop-into-coca-cola) [Powered By World V2-Preview](https://nosible.com/blog/announcing-atlas-the-point-in-time-knowledge-graph#powered-by-world-v2-preview) [Use Cases](https://nosible.com/blog/announcing-atlas-the-point-in-time-knowledge-graph#use-cases) [1. Mapping Supply Chain Networks](https://nosible.com/blog/announcing-atlas-the-point-in-time-knowledge-graph#1-mapping-supply-chain-networks) [2. Grounding Point-in-time LLMs](https://nosible.com/blog/announcing-atlas-the-point-in-time-knowledge-graph#2-grounding-point-in-time-llms) [3. Shrinking Covariance Matrices](https://nosible.com/blog/announcing-atlas-the-point-in-time-knowledge-graph#3-shrinking-covariance-matrices) [4. Mapping Text to Tickers](https://nosible.com/blog/announcing-atlas-the-point-in-time-knowledge-graph#4-mapping-text-to-tickers) [Sneak Peak at the Upcoming API](https://nosible.com/blog/announcing-atlas-the-point-in-time-knowledge-graph#sneak-peak-at-the-upcoming-api) [Start Your Free Trial Today](https://nosible.com/blog/announcing-atlas-the-point-in-time-knowledge-graph#start-your-free-trial-today) ## Let's hop into Coca-Cola Coca-Cola alone connects directly to 1,510 other entities through 4,269 distinct relationships in Atlas. Expand just two of those companies, Coca-Cola India and Costa Coffee, and the subgraph grows to 1,796 entities and 5,456 relationships. That includes 867 organizations, 257 products, 403 people, 56 locations, and 213 geopolitical entities. These counts cover the historical graph, with repeated observations of the same relationship counted once and alias relationships excluded. We can start with the products Coca-Cola sells, then add its subsidiaries. Now follow those subsidiaries to their products. Coca-Cola India connects us to Thums Up and Maaza. Costa Coffee connects us to Mocha Italia Signature Blend. We have gone two hops: from the parent company, through a subsidiary, to a product. ![Atlas’s Coca-Cola connections expand into a pyramid: 111 product-linked nodes and 54 subsidiary-linked nodes at the first hop, followed by the products, people, regions, and countries connected through Coca-Cola India and Costa Coffee.](https://nosible.com/images/2026/09/nosible-atlas-coca-cola-1820x780.png) The same expansion connects Coca-Cola India to [Sanket Ray](https://www.coca-colacompany.com/about-us/leadership/sanket-ray), who leads India and Southwest Asia, and to the [Southwest Asia region](https://bwretailworld.com/news/coca-cola-india-appoints-irene-tan-as-vp-of-human-resources-for-india-southwest-asia) itself. Costa Coffee connects to the UK through its operations. So we can move from a parent company to the people running its businesses, the products they sell, and the markets they operate in. These relationships answer different questions. Who is responsible for growth in India? Which brands sit within that business? Which operations could be affected by a change in UK policy? Atlas connects the people, products, companies, locations, and jurisdictions needed to investigate those questions. Your agents can follow each connection back to its supporting evidence. ## Powered By World V2-Preview In addition to announcing the release of NOSIBLE Atlas, we are pleased to announce the release of World V2-Preview. World V1.2 already described events, identified entities, mapped tickers, classified events across 15 taxonomies, scored market signals, tracked coverage and timing, assigned geographic coordinates, supported semantic search, and maintained detailed provenance. V2-Preview takes World to the next level. In addition to everything World V1.2 did, it adds English translations, Atlas-powered ticker mapping, better event ontology classifications, better named entity recognition, entity relationship extraction, and geocoding to multiple locations. Take Coca-Cola → Sprite above. The event describes Coca-Cola's plans to grow its billion-dollar brands in India. Here is the complete V2-Preview record. JSON / triples Copy JSON Explore fields triples NEW tickers NEW english NEW event.points NEW entities NEW locations NEW availability NEW event_id version has_tickers event coverage signals provenance oai_vector ``` { "triples": [ { "subject": "Coca-Cola", "subject_type": "ORG", "predicate": "operates_in", "object": "India", "object_type": "GPE", "original_score": 0.99951171875, "reranker_score": 0.9998751878738404 }, { "subject": "Sprite", "subject_type": "PRODUCT", "predicate": "brand_of", "object": "Coca-Cola", "object_type": "ORG", "original_score": 0.98681640625, "reranker_score": 0.999619960784912 }, { "subject": "Henrique Braun", "subject_type": "PERSON", "predicate": "employed_by", "object": "Coca-Cola", "object_type": "ORG", "original_score": 0.99365234375, "reranker_score": 0.9992985725402832 }, { "subject": "Coca-Cola", "subject_type": "ORG", "predicate": "owns", "object": "Sprite", "object_type": "PRODUCT", "original_score": 0.9990234375, "reranker_score": 0.9974836707115172 }, { "subject": "Henrique Braun", "subject_type": "PERSON", "predicate": "holds_position_at", "object": "Coca-Cola", "object_type": "ORG", "original_score": 0.99609375, "reranker_score": 0.9970130920410156 }, { "subject": "Coca-Cola", "subject_type": "ORG", "predicate": "produces_product", "object": "Sprite", "object_type": "PRODUCT", "original_score": 0.87646484375, "reranker_score": 0.99014014005661 }, { "subject": "Thums Up", "subject_type": "PRODUCT", "predicate": "brand_of", "object": "Coca-Cola", "object_type": "ORG", "original_score": 0.986328125, "reranker_score": 0.98967182636261 }, { "subject": "Coca-Cola", "subject_type": "ORG", "predicate": "owns", "object": "Thums Up", "object_type": "PRODUCT", "original_score": 0.9990234375, "reranker_score": 0.9760608077049256 }, { "subject": "Coca-Cola", "subject_type": "ORG", "predicate": "sells_product", "object": "Sprite", "object_type": "PRODUCT", "original_score": 0.92919921875, "reranker_score": 0.9676862359046936 }, { "subject": "Coca-Cola", "subject_type": "ORG", "predicate": "develops_product", "object": "Sprite", "object_type": "PRODUCT", "original_score": 0.60693359375, "reranker_score": 0.96518474817276 }, { "subject": "Coca-Cola", "subject_type": "ORG", "predicate": "produces_product", "object": "Thums Up", "object_type": "PRODUCT", "original_score": 0.88671875, "reranker_score": 0.9585376977920532 }, { "subject": "Coca-Cola", "subject_type": "ORG", "predicate": "supplies_product", "object": "Sprite", "object_type": "PRODUCT", "original_score": 0.63671875, "reranker_score": 0.9530133008956908 }, { "subject": "Sprite", "subject_type": "PRODUCT", "predicate": "available_in", "object": "India", "object_type": "GPE", "original_score": 0.99853515625, "reranker_score": 0.9440738558769226 }, { "subject": "Thums Up", "subject_type": "PRODUCT", "predicate": "available_in", "object": "India", "object_type": "GPE", "original_score": 0.99853515625, "reranker_score": 0.9124361872673036 }, { "subject": "Coca-Cola", "subject_type": "ORG", "predicate": "supplies_product", "object": "Thums Up", "object_type": "PRODUCT", "original_score": 0.62646484375, "reranker_score": 0.901704490184784 }, { "subject": "Coca-Cola", "subject_type": "ORG", "predicate": "sells_product", "object": "Thums Up", "object_type": "PRODUCT", "original_score": 0.91650390625, "reranker_score": 0.8835880160331726 }, { "subject": "Coca-Cola", "subject_type": "ORG", "predicate": "invests_in", "object": "India", "object_type": "GPE", "original_score": 0.9736328125, "reranker_score": 0.8790948390960693 }, { "subject": "Thums Up", "subject_type": "PRODUCT", "predicate": "located_in", "object": "India", "object_type": "GPE", "original_score": 0.87158203125, "reranker_score": 0.8574699759483337 }, { "subject": "Coca-Cola", "subject_type": "ORG", "predicate": "planned_investment_in", "object": "India", "object_type": "GPE", "original_score": 0.9970703125, "reranker_score": 0.8543385863304138 }, { "subject": "Coca-Cola", "subject_type": "ORG", "predicate": "distributes_product", "object": "Sprite", "object_type": "PRODUCT", "original_score": 0.97412109375, "reranker_score": 0.7562676072120667 }, { "subject": "Henrique Braun", "subject_type": "PERSON", "predicate": "publicly_supports", "object": "Coca-Cola", "object_type": "ORG", "original_score": 0.7763671875, "reranker_score": 0.6935767531394958 } ] } ``` 193 lines To elaborate on the new features: | Feature | What changed | | --- | --- | | **Foreign Event English Translations** | We use Google's [TranslateGemma](https://huggingface.co/google/translategemma-4b-it) to translate titles, descriptions, and factual points into English. Original text stays in `event`; translations sit in `english`. Your models can analyze news across languages, with the original text available to check against each translation. | | **Multi-hop Ticker Mapping with Atlas** | We use personalized PageRank on a constrained Atlas graph to connect products, people, and subsidiaries to companies. Rankings reward company-specific connections. Funds can map text to candidate tickers, with a relationship path they can inspect for each mapping. | | **Better Ontology Classifications** | Our classifier combines Google's [EmbeddingGemma](https://huggingface.co/google/embeddinggemma-300m) with trained heads for 15 taxonomies. It assigns probabilities to event categories instead of relying on embedding similarity. You receive ranked classifications to filter news events by category and confidence. | | **Better Named Entity Recognition** | We fine-tuned [GLiNER V2.5](https://huggingface.co/gliner-community/gliner_large-v2.5) to identify people, organizations, products, locations, and geopolitical entities in event text. We then normalize their names. Each entity carries a type, mention count, and confidence score, so you can match names and filter uncertain detections. | | **Entity Relationship Extraction** | We fine-tuned [GLiNER](https://huggingface.co/knowledgator/gliner-relex-large-v0.5) to extract relationships and built a proprietary edge reranker to check the supporting text. Each connection carries scores from both models. Atlas joins these relationships across events, giving your agents a network of facts they can trace back to the underlying evidence. | | **Geocoding events to multiple locations** | V2-Preview maps events to multiple locations, capturing the countries and regions involved in a cross-border development. Your models can connect the same event to several markets, helping you investigate portfolio exposure without reducing the event to a single coordinate. | ## Use Cases ### 1. Mapping Supply Chain Networks Businesses depend on each other. Suppliers provide inputs. Customers provide revenue. Models need to understand both. A trade restriction affecting one business can change production, costs, and earnings at others. [Research by Acemoglu and Tahbaz-Salehi](https://www.nber.org/papers/w27565) explains how disruptions spread through these networks. Take a manufacturer that buys electronic components from a domestic supplier. That supplier imports a critical input from a company facing export restrictions. The manufacturer may have no direct relationship with the restricted company, but it still needs the component. How much inventory remains? Can the input be replaced? What happens to production if that input cannot be replaced? Trade restrictions can reach companies through indirect suppliers. These relationships matter for investment research because markets do not always price them immediately. [Cohen and Frazzini found](https://pages.stern.nyu.edu/~afrazzin/pdf/Economic%20Links%20and%20Predictable%20Returns%20-%20Cohen%20and%20Frazzini.pdf) that news about customers was reflected in suppliers’ stock prices with a delay. The connections were publicly available. Investors were not fully accounting for them. Atlas lets customers investigate those connections across their portfolios. Which holdings depend on the same supplier? Which customers face trade restrictions? Who owns the companies involved? Internal procurement records can add purchase volumes, inventory, and alternative suppliers to quantify the exposure. Take China’s April 2025 export restrictions on rare earths and magnets. By June, shortages had disrupted [Ford production in Chicago](https://think.ing.com/articles/chinas-crackdown-on-rare-earth-causes-alarm-automotive-industry/) and [Suzuki Swift production at Sagara in Japan](https://www.moneycontrol.com/news/business/suzuki-motor-halted-swift-production-due-to-china-s-rare-earth-curb-sources-say-13099201.html). Material restrictions had disrupted vehicle production. Which carmakers depended on those inputs? Which suppliers connected them? This is where ownership and supply relationships become useful together. Customers can use Atlas to trace the suppliers, customers, and owners connected to the restriction, using relationships known when it was announced. Which portfolio holdings were exposed? Compare those findings with later company disclosures. Did the expected impacts occur? Could that knowledge have helped revise earnings forecasts or adjust portfolio positions earlier? ### 2. Grounding Point-in-time LLMs Everybody is talking about “grounding” these days. What is it and why does it matter? Grounding an AI means giving it evidence that it can use to solve a problem. It’s important because AIs can give convincing answers based on incorrect, outdated, or incomplete information. Grounding gives AIs relevant and timely facts to work with and gives humans sources to check. Knowledge graphs are useful for this because they connect facts. An AI researching a company can follow its ownership, investigate its suppliers, and retrieve the evidence behind those relationships. [Microsoft Research demonstrated this with GraphRAG](https://www.microsoft.com/en-us/research/blog/graphrag-new-tool-for-complex-data-discovery-now-on-github/), which produced more comprehensive answers to questions requiring information from across an entire collection of documents. Microsoft asked both systems the same questions about news articles. An LLM judge compared their answers. GraphRAG gave more complete answers more often than standard RAG, which retrieves matching passages from documents. **GraphRAG vs. standard RAG: which gave the better answer?** | What the judge compared | Winner | How often GraphRAG won | | --- | --- | --- | | Which answer covered more of the question? | **GraphRAG** | **72% of comparisons** | | Which answer included more varied perspectives? | **GraphRAG** | **62% of comparisons** | Source: [Microsoft’s GraphRAG paper, Figure 4 and Section 3.6](https://arxiv.org/html/2404.16130v1#S3.SS6). News dataset, root-level graph summaries, 125 questions with five evaluations each. These percentages measure preference, not accuracy. Standard RAG was better at giving direct answers. And the best part? GraphRAG massively reduced token consumption. **Context tokens per question: source documents vs. GraphRAG summaries** | Dataset | Summarizing source documents | Using GraphRAG summaries | Reduction | | --- | --- | --- | --- | | News articles | 1,707,694 tokens | 39,770 tokens | **97.7%** | | Podcast transcripts | 1,014,611 tokens | 26,657 tokens | **97.4%** | Source: [Microsoft’s GraphRAG paper, Table 3](https://arxiv.org/html/2404.16130v1#S3.SS6). Highest-level graph summaries compared with summarizing the source collection, rather than standard retrieval. Figures exclude the initial cost of building the graph. But what if you want the AI to answer a question as it would have in 2022??? Now we have two problems. The evidence must have been available in 2022, and the AI must not know what happened afterwards. Giving a modern LLM old documents only addresses the first problem. It may still already know the outcome a priori from its training data. This is why researchers are training point-in-time LLMs. [ChronoLLM](https://chronollm.com/), for example, offers model vintages designed around historical knowledge cutoffs. Atlas could provide the corroborated evidence available at each historical research date. Together, they would let us test investment decisions generated using a point-in-time model that was grounded and enhanced by a point-in-time knowledge graph. Talk about alpha right! Take Microsoft’s January 18, 2022 offer to acquire Activision Blizzard for $95 per share. We could ask a point-in-time model whether it would invest, then use Atlas to ground its reasoning in the evidence available that day. What could go wrong? Was the potential return worth the risk? Then test its decision against the outcome. ### 3. Shrinking Covariance Matrices Portfolio construction depends on knowing how investments move together. A covariance matrix measures those relationships. Quant funds use it to estimate portfolio risk and decide how much to invest in each company. The problem is that these estimates are noisy and need to be stabilized before an optimizer uses them. Take a portfolio of 1,000 stocks. There are 499,500 distinct pairs, but one year gives us roughly 252 daily returns per stock. Underestimate how closely two stocks move together and the optimizer may allocate too much capital to both. Overestimate it and the optimizer may reject useful diversification. Either way, errors in the matrix become errors in individual position sizing decisions. Covariance shrinkage helps by blending the historical estimate with a structured model, called a shrinkage target. One approach estimates each stock’s sensitivity to the overall market and uses those sensitivities to model how stocks move together. This gives the optimizer a more stable estimate when price history is limited. But prices aren't our only source of information about companies. [Bloomberg researchers tested](https://aclanthology.org/2023.findings-emnlp.668/) whether company descriptions and knowledge graphs could improve shrinkage targets. Their graph included suppliers, subsidiaries, board members, and other relationships (Atlas has considerably more). They used this information to measure company similarities, then combined it with historical returns. The results were encouraging. Financial-graph shrinkage reduced average covariance-estimation loss by 6.73%. The strongest price-only shrinkage benchmark managed 2.59%. This means company relationships improved the risk estimate beyond what the strongest price-only benchmark achieved in this experiment. | Shrinkage target | Reduction in estimation loss | | --- | --- | | Constant variance | 2.59% | | Constant correlation | 2.22% | | Single market factor | 1.43% | | Company descriptions | 3.64% | | Wikidata knowledge graph | 4.47% | | **Bloomberg knowledge graph** | **6.73%** | Adapted from [Becquin and Esmeir, Figure 4](https://aclanthology.org/2023.findings-emnlp.668.pdf#page=7). S&P 500, 2007–2023. All six rows use shrinkage. Figures show improvement over sample covariance, evaluated against the following month’s sample covariance. Higher is better; these are risk-estimation results, not investment returns. But there was a gap in Bloomberg's research: their company knowledge came from 2021 and 2022. Using it in earlier periods would introduce knowledge of relationships that did not yet exist at that point-in-time. The authors acknowledged this limitation: “potentially leading to data leakage for the 2007-2023 period backtest.” Atlas gives customers dated relationships for testing the same approach. At each historical date, they could build a shrinkage target from the company knowledge available then and combine it with recent returns. Target construction and model fitting would follow the same cutoff. In theory you could update the target daily. Customers can compare the resulting portfolios with their existing approach. Were risk forecasts more accurate? Were allocations more stable? Did realized portfolio risk improve? Bloomberg showed that company knowledge can improve covariance estimates. Atlas lets customers test that approach using what was known at the time. ### 4. Mapping Text to Tickers A lot of useful information never mentions a listed company by name. An article might discuss the F-35, a review might mention Siri, or a filing might name a subsidiary. Funds need to connect those mentions to the companies they invest in. Searching for company names and ticker symbols only gets you so far. Atlas lets us follow the relationships instead. The [F-35](https://www.lockheedmartin.com/en-us/products/f-35.html) connects to Lockheed Martin. Siri connects to Apple. A person can connect through an employer, then through that employer’s parent company. This gives funds a way to map news, transcripts, research, and other textual datasets onto their investment universe, even when the company itself is never named. But which connections actually matter? We used [personalized PageRank](https://arxiv.org/abs/1512.04633) to rank the graph nodes connected to each company. Think of it as repeatedly following relationships from that company, returning to the starting company 25% of the time. [Relationships receive different weights](https://networkx.org/documentation/stable/reference/algorithms/generated/networkx.algorithms.link_analysis.pagerank_alg.pagerank.html) based on their type, supporting evidence, and recency. The algorithm uses all eligible connections to calculate the ranking. We also constrained the graph. This run includes people, organizations, and products connected through commercial relationships. Each relationship must meet evidence-count and age requirements. Routes into a company are limited to two connections, with the second allowed only through an organization linked by an ownership or identity relationship. Traversal stops at another company’s entry point. Finally, we adjust for specificity. A node connected to one company receives more weight than one spanning several companies. That helps distinguish a company’s own products and subsidiaries from names that appear throughout the market. Here are the ten highest-ranked connections for each of twelve companies from our September 26, 2026 run. Select a company to inspect the names, scores, and relationship chains, including the earliest evidence behind each connection. ATLAS / Eli Lilly Copy chains ↗ Explore companies Eli Lilly 10 Lockheed Martin 10 Toyota 10 Disney 10 Northrop Grumman 10 SpaceX 10 Apple 10 Netflix 10 Microsoft 10 Walmart 10 Exxon Mobil 10 JPMorgan Chase 10 Snapshot: 26 September 2026. Dates show earliest evidence, not relationship start dates. Scores are rankings, not probabilities. † Flags a historical or incorrect assertion; focus or hover for details. 1. Tirzepatide (PRODUCT) 0.076927 Tirzepatide ←develops_product— Lilly Earliest evidence: 2024-06-25 2. Mounjaro (PRODUCT) 0.032661 Mounjaro —brand_of→ Eli Lilly and Company Earliest evidence: 2023-08-22 3. Omvoh (PRODUCT) 0.017833 Omvoh —brand_of→ Eli Lilly and Company Earliest evidence: 2023-10-26 4. Zepbound (PRODUCT) 0.017152 Zepbound —brand_of→ Eli Lilly and Company Earliest evidence: 2024-04-07 5. Kisunla (PRODUCT) 0.014615 Kisunla ←produces_product— Eli Lilly and Company Earliest evidence: 2024-07-17 6. Mirikizumab (PRODUCT) 0.013906 Mirikizumab —brand_of→ Lilly Earliest evidence: 2024-09-04 7. Jaypirca (PRODUCT) 0.009205 Jaypirca ←develops_product— Eli Lilly and Company Earliest evidence: 2023-08-22 8. pirtobrutinib (PRODUCT) 0.008728 pirtobrutinib —brand_of→ Lilly Earliest evidence: 2025-04-15 9. Donanemab (PRODUCT) 0.006357 Donanemab —brand_of→ Eli Lilly and Company Earliest evidence: 2021-03-13 10. TuneLab (PRODUCT) 0.005702 TuneLab ←owns— Eli Lilly and Company Earliest evidence: 2025-09-16 10 chains / Eli Lilly 12 companies · 120 candidate connections A fund could extract names from documents, retrieve ranked company connections, and resolve the companies to tickers in its security master. Now product news can reach a portfolio watchlist. Subsidiary developments can feed parent-company research. Every proposed mapping has a relationship path that the team can inspect. These are ranked candidate connections, not confirmed event-to-ticker assignments. The scores measure graph importance, not the probability that a ticker belongs on an event. This run used a September 2026 snapshot. Historical tagging must use the evidence and corporate relationships valid at each event date, and ticker accuracy still needs to be tested separately. ## Sneak Peak at the Upcoming API Atlas is available for on-premises deployments and trials today. Later this week, we plan to make it available through the [nosible.world](https://nosible.world/) web application and API as well. The API gives your agents eleven endpoints to explore Atlas. Ask a question, expand a company’s relationships, or find the connections between two entities. Use an `as_of` date to ground your research in what was known at the time. | Endpoint | What it does | | --- | --- | | `POST /graphs/prompt` | Ask in plain English and get a graph back | | `POST /graphs/entities` | Get the neighborhood around one or more entities | | `POST /graphs/paths` | Find how two entities are connected | | `POST /graphs/ownership` | Trace who ultimately owns whom | | `POST /analysis/rankings` | Rank the most important entities in a graph | | `POST /analysis/communities` | Find clusters of closely linked entities | | `POST /analysis/bridges` | Find the entities that connect otherwise separate groups | | `GET /entities/relationships` | List an entity’s relationships, strongest first | | `GET /entities/similar` | Find entities with similar connections | | `GET /relationships/events` | Retrieve the evidence behind a relationship | | `GET /relationships/types` | List the available relationship types | All paths above use the `/api/atlas` prefix. For an on-premises deployment, set `ATLAS_URL` to your server’s address and `ATLAS_KEY` to your API key. Who supplied NVIDIA at the end of 2022? Here is how you would ask: ``` curl -X POST "$ATLAS_URL/api/atlas/graphs/prompt" \ -H "Authorization: Bearer $ATLAS_KEY" \ -H "Content-Type: application/json" \ -d '{"prompt": "Who are the main suppliers of Nvidia?", "as_of": "2022-12-31"}' ``` The response gives your agent a graph of relationships it can follow back to the supporting evidence. API documentation will accompany the release later this week. ## Start Your Free Trial Today NOSIBLE is on a mission to build AIs that know everything, all the time. Atlas puts 35 million corroborated facts at your agents’ fingertips through a web-scale, point-in-time knowledge graph. Deploy it on-premises and add your proprietary intelligence to find new sources of alpha. If that excites you, [let’s talk](mailto:stuart@nosible.com?subject=Atlas%20demo). S3 delivery, local API, or web API. ## Start your 90-day trial 90-day trial Review the diligence documents, schema, sample data, and delivery options. Then start your trial. [Start trial](https://nosible.com/start-trial) [Review docs](https://docs.nosible.com/) [All Research](https://nosible.com/blog) Related Research ![NOSIBLE World knowledge graph showing entity connections over a decade](https://nosible.com/images/2026/07/kg-hero-decade.png) [Web Search](https://nosible.com/blog/tag/web-search) [Information Retrieval](https://nosible.com/blog/tag/information-retrieval) [NOSIBLE World](https://nosible.com/blog/tag/nosible-world) ### [Point-in-Time Knowledge Graphs over Named Entities with NOSIBLE World](https://nosible.com/blog/point-in-time-knowledge-graphs-over-named-entities) 2026-07-16 8 min read ![Abstract vortex illustration representing self-organizing web-scale search facets](https://nosible.com/blog/illustrations/vortex.png) [Web Search](https://nosible.com/blog/tag/web-search) [Information Retrieval](https://nosible.com/blog/tag/information-retrieval) ### [Can Faceted Search at Web-Scale Self Organize?](https://nosible.com/blog/can-faceted-search-at-web-scale-self-organize) 2025-10-16 7 min read ![Cybernaut-1 illustration representing agentic search with Monte Carlo Tree Search](https://nosible.com/blog/illustrations/cyber.png) [Web Search](https://nosible.com/blog/tag/web-search) [Information Retrieval](https://nosible.com/blog/tag/information-retrieval) [Artificial Intelligence](https://nosible.com/blog/tag/artificial-intelligence) [Machine Learning](https://nosible.com/blog/tag/machine-learning) ### [Introducing Cybernaut-1: Agentic Search using MCTS](https://nosible.com/blog/introducing-cybernaut-1-agentic-search-with-mcts) 2025-08-26 2 min read > Meet Atlas: 35 million corroborated facts for grounding AI, testing portfolio risk, and mapping supply chains with point-in-time company knowledge. **URL:** https://nosible.com/blog/announcing-atlas-the-point-in-time-knowledge-graph