feat(retrieval): second meta-belief consumer — adaptive bm25f_anchor_weight (#757) - #787
Merged
robotrocketscience merged 5 commits intoMay 14, 2026
Merged
Conversation
…ta-belief env flag First slice of #757 consumer wiring. Pure additions — no caller changes yet, so retrieval is byte-identical until the read-path commits land. Adds: - META_BM25F_ANCHOR_WEIGHT_KEY, BM25F_ANCHOR_WEIGHT_FLOOR (1), BM25F_ANCHOR_WEIGHT_CEIL (10), META_BM25F_ANCHOR_WEIGHT_STATIC_DEFAULT (0.5), META_BM25F_ANCHOR_WEIGHT_POSTERIOR_DECAY_SECONDS (30d). - ENV_META_BELIEF_BM25F_ANCHOR_WEIGHT = "AELFRICE_META_BELIEF_BM25F_ANCHOR_WEIGHT", ships default-OFF behind the #437 A/B bench-gate clause. - decode_meta_bm25f_anchor_weight(v) — log-linear interpolation [0, 1] -> [1, 10], rounded to int (BM25Index.build replicates the anchor token stream by an int count, see bm25.py:243). v=0.5 decodes to 3, matching DEFAULT_ANCHOR_WEIGHT exactly so the cold-start install preserves byte-identical retrieval order. - is_meta_belief_bm25f_anchor_weight_enabled() — truthy-token resolver mirroring is_meta_belief_half_life_enabled(), including the '=enabled' spelling from the umbrella #480 issue body. Encoding lives in the consumer per the 2026-05-13 #756 ratification. The substrate stays pattern-uniform across #480 sub-tasks B–F; each consumer picks its own bounds at integration time. Scope-down note: #757's body asked for per-field weights, but aelfrice's BM25F has one tunable (anchor_weight). Scope was ratified down to anchor_weight only before any code was written.
…a-aware resolver Second slice of #757. Still no caller change in retrieve() or _l1_hits — the new resolve_bm25f_anchor_weight_with_meta() function exists in parallel, ready for the third commit to wire in at the BM25IndexCache construction site. Adds: - install_bm25f_anchor_weight_meta_belief(store, *, now_ts) — idempotent install of the #757 row with the v3.x ratified defaults: bm25_l0_ratio signal only (relevance deferred to #779 per D4, same split as #756), 30d posterior decay, static_default=0.5 (decodes to 3, matching bm25.DEFAULT_ANCHOR_WEIGHT exactly). - resolve_bm25f_anchor_weight_with_meta(store, *, now_ts, explicit=None) — three-tier precedence: explicit kwarg > meta-belief (gated on the env flag) > bm25.DEFAULT_ANCHOR_WEIGHT. Unlike the #756 resolver this one has no env-var or TOML override layer because anchor_weight has never had a user-facing config knob — it has been a code constant since v1.5/#148 R3. Keeping the surface narrow avoids introducing a public config that we'd have to keep stable. The bm25.DEFAULT_ANCHOR_WEIGHT import is local (inside both helpers) to avoid a top-of-module aelfrice.retrieval ↔ aelfrice.bm25 cycle on the forward-ref direction — bm25 already imports nothing from retrieval, but a top-level reverse import would force ordering constraints under test collection. Smoke-checked the precedence ladder against an in-memory store: no store → 3, store + flag-off → 3, store + flag-on + cold meta-belief → 3 (byte-identical to baseline at cold start), explicit=7 always wins, re-install returns False.
…ratio signal update Closes the #757 loop in _l1_hits. The BM25F branch's BM25IndexCache construction now drives anchor_weight through resolve_bm25f_anchor_weight_with_meta when the caller hasn't supplied a cache, and on each query (when the meta feature flag is on) records a bm25_l0_ratio sub-posterior update: evidence is the fraction of store-locked beliefs that appeared in BM25F's top-K for the query. Wiring: - Caller-supplied bm25f_cache stays the unconditional override. The bench harness and unit tests still pin specific anchor_weights via the cache, so determinism contracts there are untouched. - The signal update fires inside _l1_hits because both the BM25F raw scored_pairs (top-K bid list) and the store handle are local there; threading them to retrieve_v2's tail would have been the same logic at one more level of indirection. - update_meta_belief is wrapped in try/except — store errors are printed to stderr and swallowed, mirroring the #756 latency-signal posture in retrieve_v2 (commit 521f668). Retrieval must never raise on a meta-belief write failure. - No locked beliefs in the store means no signal update (skip; evidence would be 0/0 division otherwise). That's the right fail-soft: empty L0 is uninformative for an L0-coverage metric. Default-OFF byte-identical contract: with the env flag unset, the read path collapses to BM25IndexCache(store, anchor_weight=3) via the resolver's fallback to bm25.DEFAULT_ANCHOR_WEIGHT, identical to the pre-#757 BM25IndexCache(store) call. Signal updates are fully gated on the same env flag. All 27 BM25-touching tests still pass under -k bm25.
…nsumer Mirrors tests/test_temporal_half_life_meta.py from PR #784 (#756), scoped to the smaller surface #757 exposes: - Log-linear encoding: floor / ceil / mid=DEFAULT_ANCHOR_WEIGHT parity / out-of-range clamping / int-valued return / monotonic non-decreasing over the unit interval. - Env feature flag: default-off / truthy + 'enabled' / falsy. - Constants sanity: pinned to the #757 ratified defaults so a careless edit surfaces in this file, not in a bench surprise. - Install helper: first-call True / second-call no-op (idempotency contract). - Resolver precedence: no-store fallback / flag-off ignores installed meta-belief / flag-on cold-start matches static default / explicit kwarg wins / flag-on-no-install falls through without materialising a phantom row. - _l1_hits integration: default-OFF byte-identical posture (no signal recorded with flag-off even when row is installed) / flag-on records a bm25_l0_ratio sub-posterior / flag-on no-install no-crash / caller-supplied bm25f_cache bypasses the resolver (bench harness pin contract). 30 tests, all green under uv run pytest.
|
Warning Rate limit exceeded
You’ve run out of usage credits. Purchase more in the billing tab. ⌛ How to resolve this issue?After the wait time has elapsed, a review can be triggered using the We recommend that you space out your commits to avoid hitting the rate limit. 🚦 How do rate limits work?CodeRabbit enforces hourly rate limits for each developer per organization. Our paid plans have higher rate limits than the trial, open-source and free plans. In all cases, we re-allow further reviews after a brief timeout. Please see our FAQ for further information. ℹ️ Review info⚙️ Run configurationConfiguration used: Path: .coderabbit.yaml Review profile: CHILL Plan: Pro Run ID: ⛔ Files ignored due to path filters (1)
📒 Files selected for processing (2)
✨ Finishing Touches🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
PR-size soft capThis PR is over the advisory size threshold:
Bigger PRs collide with more open work, which under the parallel-session workflow tends to produce repeated This is advisory only — nothing is blocked. If the size is intentional (large refactor, module removal, generated code), apply the |
![sourcery-ai[bot]](https://avatars.githubusercontent.com/in/48477?s=60&v=4){kind=small}
There was a problem hiding this comment.
Sorry @robotrocketscience, you have reached your weekly rate limit of 500000 diff characters.
Please try again later or upgrade to continue using Sourcery
This was referenced May 14, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Closes #757. Sub-task C of umbrella #480 — pattern reuse of #756 (just shipped via #784), scoped to the actual single-knob BM25F that aelfrice has rather than the multi-field BM25F the issue body assumed.
Scope-down call (ratified before any code)
#757's issue body assumed a per-field BM25F with weights for title/body/tags. The actual code in
src/aelfrice/bm25.pyhas one tunable:anchor_weight(DEFAULT_ANCHOR_WEIGHT = 3), the weight on the incoming-anchor token stream from #148 R3. There are no title/body/tags fields. Scope was ratified down toanchor_weightonly before any code was written; this PR honours that.Commit layout (5 atomic, all SSH-signed)
feat(retrieval): log-linear bounded bm25f_anchor_weight encoding + meta-belief env flag— constants, decoder, env-flag predicate. Pure additions, no caller change.feat(retrieval): install_bm25f_anchor_weight_meta_belief helper + meta-aware resolver— installer + 3-tier resolver. Still no caller change.feat(retrieval): wire meta-belief anchor_weight into BM25F + bm25_l0_ratio signal update—_l1_hitsBM25F branch reads through the resolver when no cache is supplied; emitsbm25_l0_ratioevidence per query when the flag is on.test(retrieval): coverage for #757 bm25f_anchor_weight meta-belief consumer— 30 new tests mirroringtests/test_temporal_half_life_meta.py.docs(changelog): note adaptive bm25f_anchor_weight consumer (#757)— Unreleased entry.Default-OFF byte-identical contract
Without
AELFRICE_META_BELIEF_BM25F_ANCHOR_WEIGHT, the read path collapses toBM25IndexCache(store, anchor_weight=3)via the resolver's fallback tobm25.DEFAULT_ANCHOR_WEIGHT— identical to the pre-#757BM25IndexCache(store)call. Signal updates are gated on the same flag. Caller-suppliedbm25f_cacheis honoured unconditionally so the bench harness pin contract is preserved.Cold-start parity
static_default = 0.5decodes throughdecode_meta_bm25f_anchor_weighttoround(sqrt(1*10)) = round(3.162) = 3 = DEFAULT_ANCHOR_WEIGHT. Installing the meta-belief preserves byte-identical retrieval order until evidence actually moves the posterior.Signal
bm25_l0_ratio= fraction of store-locked beliefs that BM25F's top-K also surfaced. Updated per query when both the flag is on and the store has at least one locked belief. Empty L0 → skip (undefined metric). Per the same D4 split #756 followed, therelevancesignal stays deferred to #779.Test plan
uv run pytest tests/test_bm25f_anchor_weight_meta.py)tests/test_bm25_index.pystill passesbenchmarks/results/v2.0.0.jsonis canonical,uv sync && aelf bench all#437 A/B corpus) — gates flipping the env flag on by default; outside this PR