From Vector Search to Weighted Backtrace Graphs

Vector search answers a narrow question well: which records are nearest to this query in embedding space? For a lot of retrieval-augmented generation, that is enough. For high-stakes answer generation — where a wrong citation is a liability, not an inconvenience — nearest-neighbor ranking alone leaves too much unsaid.

WDBX, the Weighted Directed Backtrace eXecution engine, is our answer to what ranking leaves out. It keeps retrieval context as weighted paths so a system can inspect why a result was produced, which sources were used, and where confidence dropped.

What ranking cannot tell you

A ranked list gives you the top-k records and a similarity score. It does not tell you whether two of those records contradict each other, whether the third was only retrieved because it shares boilerplate with the query, or whether the answer actually depended on the record ranked seventh. In an audit, "the model saw these ten documents" is not the same claim as "the answer rests on these two."

Backtrace graphs make the dependency explicit. Each retrieved record is a node; edges carry weights that reflect how strongly a record contributed to the generated answer, and the graph preserves the path from query to claim.

Confidence as a graph property

Once retrieval is a weighted graph rather than a flat list, confidence stops being a single scalar bolted onto the response. It becomes a property you can read off the structure: source coverage, graph distance from query to supporting evidence, and contradiction between paths all become measurable signals. Combining them gives operators a confidence band they can actually reason about instead of a number with no provenance.

Durable, inspectable, portable

WDBX persists these structures with JSONL snapshots and integrity checks, so the graph that produced an answer last quarter can be reopened and walked today. The store exposes key-value, vector (cosine search with a SIMD path), and block/spatial surfaces, and disabled builds fail closed with explicit errors rather than silently degrading.

The research bet is simple: as agents take on higher-stakes work, the retrieval layer has to carry relationships, not just neighbors. Ranking tells you what was close. Backtrace tells you what was responsible.