DASH

Evidence-first vector database for citation-grade RAG.

Status

Production-Ready v0.1

Repository

GitHub →

Install

npm install @sidianlabs/dash

001

THE PROBLEM

Similarity is not
accuracy.

Vector Blindness

Vector search returns the chunk closest to the query - a contradicted fact ranks exactly as high as a correct one.

Lost Provenance

Chunks carry no stance, no source quality, and no way to cite exactly where an answer came from.

Temporal Expiry

A revenue figure from Q2 keeps ranking in Q4. Standard vector stores have no concept of time or validity windows.

002

THE SOLUTION

Three Primitives

Claim

The atomic assertion. Never a blob - one fact, one confidence, one validity window.

claim_id, tenant_id
canonical_text
confidence
valid_from / valid_to

Evidence

The provenance. Where the claim came from, and which way the source leans.

evidence_id, claim_id
source_id
stance: supports | contradicts
source_quality

Edge

The relationship between two claims - typed, weighted, and explainable.

from_claim_id -> to_claim_id
relation: duplicates | refines
strength
reason_codes[]

002.5

THE COMPETITIVE LANDSCAPE

It's not just another
Vector Database.

// DASH vs The Market //

Dimension	DASH	Pinecone	Weaviate	Milvus	Qdrant	Chroma
Claim + Evidence Model	first-class	no	no	no	no	no
Contradiction Handling	first-class	no	manual	no	no	no
Temporal Validity Windows	first-class	metadata	manual	manual	manual	manual
Hash-chained Audit Log	yes	no	no	no	no	no
OpenAI /v1/embeddings Match	yes	limited	yes	proxy	proxy	yes

003

ARCHITECTURE

Evidence Memory
Engine (EME).

// Distributed Retrieval Pipeline //

01 // INGESTION

Extraction & Normalization

Claim Log (WAL)

02 // INDEXING

Mutable Delta Index

Background Compaction

Immutable Segments

03 // RETRIEVAL

Query Planner

Evidence Graph Assembler

Ranking & Calibration

RAG RETRIEVAL API

Zero
Migration.

Point any OpenAI SDK at DASH with a single environment variable. Every embeddings call site you already have keeps working, no code changes.

Python Example

# your existing code, untouched
import openai
client = openai.OpenAI(
    base_url="http://localhost:8080/v1",
    api_key="not_needed"
)

resp = client.embeddings.create(
    input="hello world",
    model="text-embedding-3-small"
)
# now backed by an auditable,
# hash-chained evidence store

004

TECHNICAL SPECIFICATION

Engineered for
Production.

// Core Capabilities //

CAPABILITY 01

Claim + Evidence + Edge data model with provenance

CAPABILITY 02

Contradiction handling and temporal validity windows

CAPABILITY 03

Drop-in OpenAI-compatible /v1/embeddings endpoint

CAPABILITY 04

Built-in HNSW ANN retrieval with durable WAL

CAPABILITY 05

Hash-chained audit log and per-tenant rate limiting

CAPABILITY 06

JWT auth, scoped API keys, and SDKs for Python, Go, and TypeScript

005

DEPLOYMENT

Use Cases

What Engineers Are Building

DASH replaces standard vector stores when accuracy, provenance, and auditability are non-negotiable.

Build RAG systems that cite sources and expose contradictions

Replace OpenAI embeddings with an auditable local endpoint

Retrieve claims within a temporal validity window

Run multi-tenant knowledge stores with scoped keys and rate limits