v1.0.0 — free-path evaluation program + identity + failure injection

production-rag v1.0.0

Clone-only free-path flagship. Still $0. Still no Vercel. Never cite production-rag.vercel.app (Ipsura).

What shipped

• Free-path eval program n=52 (golden-free-path.jsonl) with slices answerable / unanswerable / filter / hybrid-vs-dense
• Difficulty predicates: CI fails if a ranking slice is all trivial rank-1
• Scorecard HTML+JSON ablation dense/sparse/hybrid/hybrid_rerank labeled plumbing, billed=false
• HTTP transcripts: grounded, refuse, title=Filtering, stream; failure path with CACHE_ENABLED=false
• /v1/ready collection identity: embedder_id, chunker_version, doc_count, corpus_hash
• Cache key includes corpus identity (no cross-hit); typed wrong_collection
• Provider/store failures → 503 with refused: false (never soft refuse)
• Local load harness: docs/assets/load.json (n=100, honest single-process note)
• ADR-0014 corpus hash in cache key; CASESTUDY ≥1500 words

Still PLANNED

• Auth / rate limit (P5)
• Hosted Qdrant / multi-worker Redis cache
• Hosted-provider quality baseline (billed run with named providers)
• Public P1 host (will not ship; clone-only is intentional)

Verify

pip install -e ".[dev]"
pytest -q
python scripts/write_compose_transcripts.py
python scripts/run_load_local.py

v0.3.0 — scorecard replay, ui-stream, CASESTUDY

production-rag v0.3.0

Builds on v0.2.0 (stream + filter-aware cache + scorecard). Still clone-only, $0 free path, no hosted URL.

See it (Week 2)

• UI stills regenerated with Stream the draft toggle visible
• New docs/assets/ui-stream.png (grounded + stream checked)
• capture_ui.py rebuilds API image; CACHE_ENABLED=false for service-failure still
• DEMO-DAY beats: filter chip title=Filtering, stream, cache hit, /evals

Scorecard is a loop (Week 3)

• Replay free-path golden items via /v1/query and /v1/query/stream
• Tests: unanswerable refuses; grounded keeps markers; cache hit on second replay when enabled
• CASESTUDY: why stream is additive; why cache keys include filters

Main: after merge of PR #4
PR: #4

v0.2.0 — stream, filter-aware cache, scorecard

production-rag v0.2.0

Clone-only free path. No hosted URL. $0 with fake providers.

Stream is additive

• POST /v1/query remains the contract
• POST /v1/query/stream is extra SSE: provisional deltas, then one terminal result (grounded or refused)
• UI stream toggle; draft is never the answer

Cache off by default

• In-process result cache; enable with CACHE_ENABLED / config for local demos
• Cache key includes filters (filter mismatch cannot reuse an unfiltered hit)

Scorecard honesty

• docs/assets/scorecard.html + GET /evals
• Free-path fixture; billed=false; not a quality claim
• DEMO-DAY filter chip is title=Filtering (source=sample does not narrow the sample corpus)

Clone-only / $0

• Qdrant local; no Vercel; auth and rate-limit stay in P5
• CI empty keys; free path stays credential-free

Main: aff1674
PR: #3

production-rag v0.1.0 — free-path flagship

production-rag v0.1.0 — free-path flagship

First tagged release of the flagship RAG service. A clean clone can exercise ingest,
retrieval, answering, the web UI, and both evaluation tiers on deterministic local
providers: no credential, no billed call, and no signup.

Target commit: 678c5543baf0f4a723dc823de1f19162ba54b4a9

What shipped

• Hybrid retrieval with RRF: dense and sparse/BM25 branches run in Qdrant and are
  fused by rank rather than by incomparable raw scores.
• Optional fail-open reranking: a cross-encoder can reorder the fused shortlist; a
  reranker failure is reported and falls back to fusion order instead of failing the query.
• Grounded generation: the LangGraph query path resolves [n] markers against the
  exact prompt context, removes invalid markers, and explicitly refuses when evidence is
  insufficient.
• Two offline evaluation tiers: retrieval and answer/citation behaviour are measured
  separately, with provenance, paired statistics, and generated documentation checks.
• A reproducible free demo: one script starts Qdrant, ingests the sample corpus, and
  serves the query UI with local providers pinned.
• Credential-free CI: lint, strict type checking, tests, ingest, both eval tiers, and
  an HTTP query run with provider keys set to empty values.
• Production-facing seams: request IDs, structured logs, readiness, timings, optional
  tracing, runbooks, ADRs, security guidance, and an explicit ship-status page.

Honesty boundary

• The published scorecard uses deterministic fake providers. It validates plumbing,
  contracts, provenance, and publication; it does not measure hosted retrieval or answer
  quality.
• Authentication, authorization, rate limiting, a metrics endpoint, production retry and
  circuit-breaker policy, multi-tenancy, and load/concurrency claims are not included.
• There is no hosted demo and no attached binary. The supported demo is the reproducible
  local path documented in the README.

Start here

• Try it free
• Ship notes
• Engineering case study
• Changelog