Ported from the repository CHANGELOG. The source of truth lives with the code.

Changelog

All notable changes to DASH are documented here. The format follows Keep a Changelog and the project adheres to Semantic Versioning.

[Unreleased]

Added

• OpenAI-compatible /v1/embeddings endpoint on the retrieval service. Any OpenAI client (langchain, llama-index, semantic-kernel, the openai Python SDK) can point at DASH with one environment variable (OPENAI_API_BASE=http://localhost:8080/v1). 17 tests cover the wire-format compatibility, error envelope, and HTTP-level integration.
• Python SDK dash-py - idiomatic Python client with sync + async, typed dataclasses, OpenAI drop-in examples, RAG example showing the Claim + Evidence + Contradiction differentiator. 59 tests pass.
• Go SDK dash-go - Go 1.21+ module, idiomatic functional options
  • service pattern, full OpenAI drop-in compatibility, errors.Is/As support. 86 tests pass.
• TypeScript SDK dash-ts - ESM-first, Node 18+, zero runtime deps, full type safety with discriminated unions, drop-in OpenAI compatibility. 65 tests pass.
• Docker + docker-compose - multi-stage, multi-arch (linux/amd64, linux/arm64), non-root runtime, healthcheck, dependency ordering, dev overlay with hot-reload. docker compose up -d brings the ingestion + retrieval services online.
• Quickstart + README + comparison docs - docs/quickstart.md (5-minute path from clone to first query), README.md (rewrite with hero section and the differentiator showcase), docs/comparison.md (DASH vs Pinecone/Weaviate/Milvus/Qdrant/Chroma).
• Real embedding integration - env-driven provider selection (DASH_EMBEDDING_PROVIDER=hash|ollama|openai) so deployments can wire up real semantic embeddings instead of the hash-based default.
• Semantic-first retrieval - new InMemoryStore::retrieve_semantic method. When the caller passes a pre-computed query vector, the dense-similarity score becomes the primary ranking signal (cosine in [-1, 1], mapped to [0, 1]); the lexical/BM25 score becomes a small tie-breaker. 3 new integration tests cover the semantic-first guarantee.
• redb persistence (PR 1, additive, default off) - DiskBackedStore struct in pkg/store/src/disk.rs provides on-disk durability for claims, evidence, edges, vectors, and the tenant→claim set. Enabled via DASH_INGEST_PERSISTENCE_PATH / DASH_RETRIEVAL_PERSISTENCE_PATH env vars. If unset (the default), DASH runs in WAL-only mode - the pre-redb behavior is preserved bit-for-bit. 5 new integration tests cover round-trip, fallback, tenant consistency, ANN rebuild, and open-failure paths. The design doc at docs/plans/2026-06-13-redb-persistence-design.md describes the full 3-PR phasing.
• cargo-fuzz harnesses - fuzz/ directory with 4 focused targets (fuzz_jwt, fuzz_openai_embeddings, fuzz_ranking, fuzz_wal_parse) that exercise the JWT verifier, OpenAI request parser, ranking function, and WAL record parser against arbitrary input. Run via cargo +nightly fuzz run <target>. See fuzz/README.md.
• Performance benchmark suite - tests/benchmarks/src/perf_bench.rs with 5 scenarios: ingest_throughput_sequential (in-memory and persistent), retrieve_throughput_lexical, retrieve_throughput_semantic, ann_search_throughput_at_scale, wal_replay_throughput. CLI flags --scenario, --iterations, --warmup. Baseline numbers published in docs/benchmarks/performance.md.
• Public evidence_for_claim accessor on InMemoryStore for test introspection of ingested evidence.
• OpenAIErrorResponse with DashError interface and DashAPIError/DashConnectionError concrete types; from_response factory tolerates both OpenAI and ad-hoc error shapes.

Changed

• JSON parsing in services/ingestion - replaced 633 lines of hand-rolled JSON parsing in transport/payload.rs with serde-based deserialization. transport/json.rs is now a thin compat shim for the test suite.
• Hand-rolled crypto in pkg/auth - replaced SHA-256, HMAC-SHA256, base64url, and the recursive-descent JSON parser with jsonwebtoken, serde_json, base64, sha2, and hex. Public API preserved. 16 tests pass (was 7 in the hand-rolled impl - added 9 new edge-case tests including a FIPS SHA-256 known vector).
• Hand-rolled HNSW in pkg/store - replaced the 4-level O(N²)-per-insert HNSW scaffolding with the usearch crate. Per-tenant ANN graphs are still isolated and dimension-pinned; the new index is substantially faster on the benchmark suite.
• Schema types in pkg/schema - added Serialize/Deserialize derives to all public domain types with #[serde(default)] on optional fields.
• API request/response types in services/ingestion/src/api.rs - added Serialize/Deserialize derives. 5 new serde round-trip tests.
• Half-built axum runtime paths removed from both services. The TransportRuntime::Axum branch that exited with an error when the async-transport feature was absent is gone; the runtime path is now a single direct call to serve_http_with_workers.
• .gitignore updated to ignore SDK build artifacts (node_modules/, dist/, __pycache__/, *.pyc, etc.).

Fixed

• All 24 prior build warnings (cfg conditions, dead code, lazy doc-continuation) resolved.
• 4 pre-existing test bugs (one in wal_persistence_and_replay_round_trip where the support_only test split Vecs and passed only one to ingest_bundle; fixed by merging into a single evidence vec).

Test counts

• Rust unit + integration tests: 379 passing (was 333 at the start of this modernization campaign; +46 new tests across schema, auth, store (unit), store (integration_retrieval), retrieval, embeddings, retrieval HTTP integration, and disk persistence).
• Python SDK: 59 tests passing.
• Go SDK: 86 tests passing.
• TypeScript SDK: 65 tests passing.
• Total across all stacks: 589 tests passing.

Known limitations

• The InMemoryStore::Clone derive was replaced with a manual impl that drops the disk handle on clone. This is a known limitation of the redb PR 1 design; the next PR will switch to Arc<DiskBackedStore> to share the handle cheaply. Cloned stores used for batched ingest staging will lose their disk attachment - this affects IngestionRuntime::ingest_batch which does self.store.clone() internally.
• The embeddings_for_claim returns the full evidence vec; for tenants with thousands of evidence per claim, this is unbounded. A future PR will add pagination.
• The performance benchmark suite does not include comparison against Pinecone/Weaviate/Milvus. Baseline numbers are internal (DASH against itself). See docs/benchmarks/performance.md for the methodology and the roadmap for the competitor-comparison work.

Earlier releases

Pre-modernization

The original EME/DASH architecture is documented in docs/architecture/eme-architecture.md (the "Evidence Memory Engine" phase 0 design). The 11-phase production rollout plan lives in docs/execution/phases/. The 2026-06-13 modernization roadmap at docs/plans/2026-06-13-dash-modernization-roadmap.md records the session-by-session deltas in detail.