DASH
Evidence-first vector database for citation-grade RAG.
DASH stores atomic claims with provenance, retrieves them with citation-grade rankings, and ships an OpenAI-compatible embeddings endpoint so any client can adopt it without changing call sites.
Why DASH
Naive RAG ranks documents by vector similarity and returns the top k chunks. That works for "summarize this article" but fails in three common enterprise cases:
- Two sources say opposite things and you have no way to demote the contradicted one.
- A fact has a temporal window and the version you retrieved is stale.
- Your auditor asks "why did the model say that" and the answer is "because a 768-dimensional number was close to a query."
DASH treats the claim - an atomic, source-bound assertion - as the primary data primitive, with evidence and citation as first-class fields on every result.
The response shape
Every retrieval response is { claim, score, supports, contradicts, citations[] }. Each citation carries its source_id, stance (supports/contradicts/neutral), source_quality, and an optional chunk_id plus span_start/span_end for character-level traceability.
The retrieval API exposes stance_mode: support_only to filter out claims that have been contradicted, and time_range: {from_unix, to_unix} to constrain results to a validity window.
This makes DASH a different kind of vector database: not the fastest pure vector index, but the most defensible one for RAG that has to ship to legal, medical, financial, and enterprise knowledge workflows.
Where to go next
- Quickstart - five minutes from clone to first cited retrieval
- Data Model - claims, evidence, edges, and stances
- Architecture - services, WAL, retrieval planner, auth
Repository
BHAWESHBHASKAR/DASH (Rust; Apache-2.0 intended).