ADR-0022: Hybrid Retrieval

Status: Accepted
Date: 2026-05-29

Context

The retrieval layer surfaced context with a single dense (bi-encoder) path: the user turn is embedded with the BGE query prefix and the nearest sub-chunks are read from Chroma (ADR-0004), then de-duplicated to parent cards (ADR-0020). Dense retrieval captures semantic similarity but misses exact lexical matches when the query and a card share rare tokens (a drug name, a device model, a specific dose unit) that the embedding smooths over. A pure lexical index has the inverse weakness: it misses paraphrase. For a medication-adherence agent whose corpus is dense with named entities, neither signal alone is sufficient.

The standard remedy is hybrid retrieval: run a lexical and a dense generator in parallel, fuse their rankings, then re-score the fused candidates with a cross-encoder that reads query and candidate jointly. This ADR records the decisions made when adding that pipeline.

Decision

Replace the dense-only retrieve step with a three-stage pipeline, gated behind a flag that defaults on and degrades gracefully to the prior behaviour.

Two parallel candidate generators over the same sub-chunk corpus: BM25 (lexical) and the existing dense Chroma path (semantic).
Reciprocal Rank Fusion combines the two ranked lists into one without score calibration between the systems.
Cross-encoder rerank re-scores the fused candidates against the query text; the survivors are then de-duplicated to parents (ADR-0020), truncated to top_k, and filtered by the existing minimum-similarity threshold.

The locked engineering choices:

(A) Reranker model. Primary BAAI/bge-reranker-v2-m3 (~568MB), state-of-the-art multilingual re-ranking. Documented fallback BAAI/bge-reranker-base (~110MB, ~3-5% nDCG@10 lower) if the primary does not fit RAM or its cold-start time is unacceptable on the deploy target. Paid rerankers (Cohere, Voyage) are out of scope here: they add an external paid dependency the demo budget excludes.
(B) BM25 library. rank-bm25 (BM25Okapi): pure Python, no compiled dependencies, MIT-licensed. Added as a main runtime dependency (~30KB) so the hybrid path is importable in every install rather than gated behind an extra.
(C) BM25 index lifecycle. Rebuilt at application startup from the same chunk list the dense store was built from; never pickled. The index is small (sub-second build) and so can never desynchronise from the dense corpus.
(D) RRF constant. A tuned Reciprocal Rank Fusion constant (the canonical Cormack et al. 2009 default), exposed as a setting for tuning.
(E) Candidate set sizes. Each generator over-fetches a bounded multiple of top_k; the reranker scores at most a bounded pool of fused candidates; the final set truncates to top_k. These candidate-pool sizes are tuned per deployment and provided as part of the engagement.
(F) Opt-in default. Hybrid is on by default; a single env flag reverts to the dense-only path for A/B comparison or recovery without a redeploy.
(G) Degradation contract. Four observable tiers via an agent.hybrid_path span attribute: full (BM25 + dense + RRF + rerank), rrf_only (reranker unavailable), dense_only (BM25 index empty), and the pre-existing refusal-on-no-match. The reranker loader returns nothing on any failure (missing files, OOM, no network on cold start) and the node drops to rrf_only rather than failing the request.
(H) Backward compatibility. The dense-only path is preserved as the documented fallback and is reachable via the opt-out flag; tests pin it.

Alternatives considered

A1: Vector-store-native hybrid (Chroma metadata filter + dense)

Use Chroma’s where filtering alongside dense search instead of a separate BM25 index.

Pro: one query path; no separate index to build.
Con: ties the hybrid semantics to one vector store; metadata filtering is not BM25 and does not rank by term frequency / inverse document frequency.
Rejected: agent-layer fusion is provider-agnostic (holds across Chroma, pgvector, etc.) and gives true BM25 ranking.

A2: Pyserini / Lucene-backed BM25

Pro: production-grade, fast at large scale.
Con: ~250MB plus a Java runtime; far past the demo footprint for a corpus this small.
Rejected on footprint and runtime-dependency grounds.

A3: Ship BM25 + dense + RRF only, skip the cross-encoder

Pro: lower per-turn latency; no 568MB model.
Con: RRF fuses ranks but cannot read query and candidate jointly; the cross-encoder is where most of the precision@k gain comes from.
Rejected for quality. The degradation contract still falls back to exactly this configuration (rrf_only) when the reranker is unavailable, so the path is exercised and supported, just not the default.

A4: Pickle the BM25 index to disk

Pro: skip the startup rebuild.
Con: adds a versioning surface that can desynchronise from the source-of-truth Chroma collection.
Rejected: the rebuild is sub-second; correctness beats a negligible startup saving.

Consequences

Positive

Recall strictly improves over dense-only for any positive-recall corpus: the fused candidate set is a superset of the dense candidates, so lexical-only matches the embedding missed are now reachable.
The cross-encoder lifts precision@k by re-scoring the fused set with full query+candidate attention.
Every degradation is observable via the agent.hybrid_path span attribute, and a request never fails merely because a model did not load.

Negative

Per-turn latency grows by the reranker inference (~50-150ms on CPU for the bounded candidate pool) plus the BM25 query (~1ms), bounded by capping the reranker input set.
The first cold start downloads the ~568MB reranker; subsequent starts use the cache. The smaller fallback model exists for footprint-constrained targets.
Existing tests that asserted on dense-only ordering or exact scores must migrate to the hybrid contract: recall-superset assertions hold, exact-order assertions do not.

Neutral

The dedupe-by-parent invariant (ADR-0020) is unchanged: it runs after fusion + rerank, still on sub-chunk identities.
Tokenization is lowercase + punctuation-strip; locale-aware tokenization for es-419 / pt-BR is deferred until recall metrics warrant it.

Implementation notes

The BM25 index wraps the rank-bm25 BM25Okapi implementation; querying returns context-chunk copies with the BM25 score set. An empty corpus yields an empty result, which is the dense_only degradation trigger.
Reciprocal rank fusion is a pure function; the fusion identity is the sub-chunk id because dedupe-by-parent runs after fusion.
The reranker wraps the sentence-transformers cross-encoder; its loader is a module-level callable that lazily imports the library, so importing the retrieval module never pulls torch. The loader returns nothing on load failure (Decision G).
Settings: retrieval_hybrid_enabled, rrf_k, reranker_model, reranker_max_input.

Future work

Locale-aware BM25 tokenization for es-419 / pt-BR if recall metrics indicate lexical misses on non-English turns.
Query expansion / HyDE / multi-query as a separate retrieval-quality step if recall@k warrants it.
Paid reranker adapters (Cohere, Voyage) behind the existing cloud extra, for deployments that opt into a managed reranking API.

Rollback

Set the hybrid opt-out env flag to restore the dense-only path with no code change; the BM25 index and reranker simply go unused. The dense path is untouched by the hybrid work and remains the fallback.