feat(bench): HyDE + RRF query expansion for low-recall categories - #213
Open
sebbsssss wants to merge 4 commits into
Open
feat(bench): HyDE + RRF query expansion for low-recall categories#213sebbsssss wants to merge 4 commits into
sebbsssss wants to merge 4 commits into
Conversation
release: promote LongMemEval 77.2% milestone — fragments + chunked + rerank
staging → main: LongMemEval-S 82.4% via Opus reader on hard cats
PMP encryption control plane → production (revoke + re-delegate)
Add --rrf-variants N flag that generates N hypothetical first-person answer statements via Sonnet 4.5 and uses each as an additional retrieval query. Result lists from all N+1 queries (original + variants) are merged via Reciprocal Rank Fusion (Cormack et al. 2009, k=60). Default-targeted categories (--rrf-categories) are SS-Pref and SS-User, which have the lowest evidence-hit rates (76.7% and 94.3% on fresh data). Closes the question-vs-document distribution gap that vector search often hits on benchmarks asking "what did I do?" where the answer-bearing memory phrases the fact directly rather than echoing the question. No effect when --rrf-variants is unset; this is opt-in for benchmarking. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
Adds opt-in HyDE (Hypothetical Document Embeddings) + Reciprocal Rank Fusion to the LongMemEval benchmark recall step. Targets the ~15 no-recall misses we identified in SS-Pref and SS-User on the fresh-data baseline.
How it works
When
--rrf-variants Nis set (default off), and the question's category matches--rrf-categories(defaultsingle-session-preference,single-session-user):cortex.recall()for each variant + the original questionRefs: HyDE (Gao et al.), RRF (Cormack et al. 2009).
Why
Vector search struggles when the question (
"Where did I redeem the coupon?") and the answer-bearing memory ("I used the coupon at Target last Tuesday") have very different surface forms. Hypothetical-answer queries close that distribution gap.Default categories chosen because they have the lowest evidence-hit rates on the current main:
Cost
Per question with
--rrf-variants 2: +1 LLM call for HyDE generation + 2 extra recall calls + 2 extra embedding calls. Roughly +30-40% cost on the targeted categories only. No effect on other categories.Test plan
npx tsx --help)cludebot-test-preview.up.railway.app(script-only change)🤖 Generated with Claude Code