We built a retrieval system with no vector database

ClawIndex uses a reasoning model to retrieve relevant information instead of cosine similarity. On HotpotQA - a multi-hop reasoning benchmark designed to stress-test retrieval systems - it scored 100% hit rate with zero hallucinations, beating FAISS at 95%. No embeddings. No chunking. No vector DB.

The problem with traditional RAG

Traditional RAG pipelines follow the same pattern: chunk the document, embed each chunk into a vector, store in a vector database, retrieve by cosine similarity, feed to the model. It's elegant in theory. In practice, it has three fatal flaws.

First, chunk size is a dead end. Small chunks retrieve precisely but lack context. Large chunks have context but retrieve poorly. No fixed size wins on both dimensions. Second, cosine similarity finds semantically similar text, not text that answers the question. Those aren't the same thing. Those aren't the same thing. Third, the retrieval step is pure math. The model has zero say in why something is retrieved.

A different approach

ClawIndex skips all of that. Instead of embedding chunks, we build a structured section index of the full document: headers, hierarchy, and enough context for a model to reason about what belongs where. Instead of cosine similarity, the model reads the index and decides what's relevant. The retrieval step is the reasoning step. One pass, not two.

We're calling this category Reasoning-Native Retrieval (RNR). The retriever isn't a lookup table. It's a reasoner.

Benchmark results

We tested both systems on the HotpotQA distractor validation set: 20 questions requiring multi-hop reasoning across multiple Wikipedia articles. FAISS missed one (the David Beckham / Manchester United multi-hop question). ClawIndex got all 20, with zero hallucinations.

The tradeoff

ClawIndex runs at ~4.8s per query. FAISS runs in milliseconds. That's a real difference and we're not pretending otherwise. But for async pipelines, batch document processing, compliance workflows, or any use case where a wrong answer costs more than a slow one: that's the right tradeoff.

Real-time chat is the one use case where FAISS wins on latency. Every other use case is a conversation.

What's next

A 100-question stress test is underway. After that: temporal awareness (RNR + staleness scoring), confidence propagation, and a multi-agent memory layer where the reasoner queries across multiple indexes simultaneously. Results will be posted here.