Into:Recall
Community-owned RAG, ranked by competition
As of · Jun 4, 10:37 UTC
A that turns retrieval-augmented generation into an open contest: nodes compete to serve the best search-and-answer pipeline, and score who actually retrieves the right thing.
What is Recall
Recall is a Bittensor subnet (SN31) built around retrieval-augmented generation, or RAG: the technique of answering a question by first fetching relevant documents and then generating a response grounded in them, with citations. According to its on-chain identity description, serve embedding models, vector search, and LLM inference, while validators independently evaluate retrieval accuracy and answer quality.
The simple version: It's like Perplexity, but the search-and-answer engine is run by a competing network of nodes instead of a single company, and the best pipeline wins the queries.
Centralized equivalent: Perplexity, or any managed RAG stack that bolts a vector database onto an LLM.
How it works:
- Miners serve the retrieval and generation pipeline: embedding models, vector search, and LLM inference, per the subnet's on-chain description.
- Validators check retrieval accuracy and answer quality, scoring which pipelines actually surface and use the right information.
Why This Matters
- The problem it solves: Language models answer from frozen training data and confidently invent things they do not know. RAG grounds answers in freshly retrieved sources with citations. Recall's stated goal is to make that retrieval-and-generation pipeline an open competition rather than a closed product.
- The opportunity: Retrieval is the backbone of most useful AI systems today. A continuously improving, citation-backed engine that anyone can plug into is a large surface to aim at.
- The Bittensor advantage: Open competition routes user queries to the top-performing pipeline and keeps pressure on every node to improve, instead of locking users to one vendor's retrieval quality.
- Traction signals: On-chain, the has positive net inflows over the past week and rose roughly 26% in that window. Beyond the on-chain identity description and market data, the slot does not list a public website, repository, or social account, so independent verification of the implementation is limited for now.
Other research from the same neighborhood of the network.
