Top AI Repos — open-source AI, indexed and scored
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
LLM powered retrieval engine designed to process a ton of sources to collect a comprehensive list of entities.
| Date | Stars |
|---|---|
| 2026-08-01 | 516 |
| 2026-08-06 | 516 |
| 2026-08-18 | 516 |
| 2026-09-06 | 517 |
| 2026-09-12 | 518 |
| 2026-09-28 | 519 |
| 2026-10-03 | 520 |
| 2026-10-05 | 519 |
| 2026-10-07 | 520 |
| 2026-10-09 | 521 |
| 2026-10-26 | 520 |
| 2026-10-31 | 521 |
| 2026-11-04 | 522 |
Today
— stars today
This week
+1 stars this week
This month
+2 stars this month
Momentum
0.0
growth rate 0.19%/day
# DeepSeek This is a new experimental architecture for a llm powered internet scale _retrieval engine_. This architecture is very different from current research agents, which are designed as _answer engines_. You can see some example results here: [https://deep-seek.vercel.app/](https://deep-seek.vercel.app/) (Note that this won't let you do real queries, since I won't be able to afford it 😅) The main difference between the 2 concepts breaks down to: - Answer engine: aggregate a ton of sources to find the one _correct_ answer. This is what "research agents" have converged on recently. Examples include [perplexity](https://perplexity.ai), [gpt-researcher](https://github.com/assafelovic/gpt-researcher), or [aomni](https://aomni.com). - Retrieval engine: process a ton of sources to collect a _comprehensive_ list of entities. So far there haven't been any serious attempts to do this yet (afaik). The end result for an answer engine is a research report, the end result for a retrieval engine is a table with all the retrieved entities and enriched columns. Here's what the end result looks like (zoomed out for scale):  This is infact a small portion of the result for this query. The actual result is so large it's impossible to screenshot. There are 94 records in the final result, which the agent gathered & enriched after browsing through 356 sources. The agent also generates a confidence score for the data in the table cells as it is enriching. Note that there are certain cells highlighted in yellow - those are cells with low confidence. Those are cases where the sources may conflict, or there are no sources at all so the agent made a best guess. This is actually a number between 0 - 1, so there can definitely be better & more creative UI to showcase the score in higher fidelity. ## Getting Started ### Installation Install any one of the following package managers - npm - yarn - pnpm - bun Follow the instructions in [Install](docs/install.md) to install the package manager and the project dependencies To run the dev server, use one of the following commands according to your package manager ```bash npm run dev # or yarn dev # or pnpm dev # or bun dev ``` Open [http://localhost:3000](http://localhost:3000) with your browser to start searching or explore the pre-built examples. Note that the examples won't actually run the agent (it costs a lot of $), it's more there to show the power and the flaws of the architecture by letting you inspect the results. If you have the environment variables set, you can run it for yourself. Note that it takes ~5 min and could cost anywhere between $0.1 - $3 worth of credits, depending on number of entities retrieved and amount of data that needs to be enriched. When running the agent, check the terminal to see logs of what's happening behind the scenes. ### Environment Variables Make sure you have API keys for [Anthropic](https://anthropic.com) and [Exa](https://exa.ai). Create a .env file, and put in the following environment variables: ``` ANTHROPIC_KEY="anthropic_api_key" EXA_KEY="exa_api_key" ``` ## Architecture The system works as a multi-step research agent. The initial user query is broken down into a plan and the answer is constructed iterately as it flows through the system. Another name for this type of architecture is [flow engineering](https://x.com/karpathy/status/1748043513156272416?s=20). The research pipeline is split into 4 main steps: 1. Plan - based on the user query, the planner constructs what the shape of the end result would look like. It does this by defining the type of entity to extract, as well as the different columns in the resulting table. The columns represent additional data that is relevant for the user's query relating to the entities. 2. Search - we use both standard keyword search and neural search to find relevant content to process (both types of search are provided by [Exa](https://exa.ai)). Keyword search
Excerpt of 6,677 characters
Read on GitHubWould you bet a product on this? Bounded 0–100 and slow moving.
matched fp:fa20aa25583ea802, topic:llm
matched fp:fa20aa25583ea802, topic:agentic
matched fp:fa20aa25583ea802, topic:knowledge-graph
