Retrieval is everything. If Tessera cannot find the right precedent for a new situation, nothing else matters. I spent this month building a hybrid retrieval system that prevents the failure modes I have seen in every other approach.
Why Single-Method Retrieval Fails
Vector similarity alone drifts. It finds conceptually related content but loses precision on specifics. A query about a specific vendor dispute returns results about disputes in general rather than the exact structural match I need.
Keyword search alone misses semantics. A search for “irreversible commitment under uncertainty” will not find the email where I wrote “once we sign this, there is no unwinding it without litigation.”
Graph traversal alone is slow and requires knowing where to start.
The Hybrid Approach
Graph traversal preserves causality. It follows the structural relationships between decisions, ensuring that retrieved precedent is connected to the current situation through meaningful decision patterns, not just surface similarity.
Vector similarity captures conceptual alignment. It finds decisions that match the meaning of the current situation even when the language is completely different.
Lexical verification (BM25) keeps facts precise. It ensures that specific identifiers, technical terms, regulatory references, and named entities are matched exactly rather than approximately.
What This Prevents
This prevents Tessera from overgeneralizing my own past. From applying the wrong lesson to the wrong context. From becoming overconfident outside validated domains. The three methods cross-check each other. If vector search suggests a precedent but graph traversal shows no structural connection, the result is downweighted. If BM25 finds an exact term match but the semantic context is wrong, it is flagged.
Tessera is deliberately conservative where I am conservative. The retrieval system enforces this by requiring convergent evidence before presenting a precedent as relevant.