Conversation
Signed-off-by: Shenjun <[email protected]>
… session recall Signed-off-by: Shenjun <[email protected]>
Signed-off-by: Shenjun <[email protected]>
Signed-off-by: Shenjun <[email protected]>
Signed-off-by: Shenjun <[email protected]>
Signed-off-by: Shenjun <[email protected]>
Signed-off-by: Shenjun <[email protected]>
Signed-off-by: Shenjun <[email protected]>
…l-write Signed-off-by: Shenjun <[email protected]>
Signed-off-by: Shenjun <[email protected]>
…and LLM fallback Implement pre-retrieval query strategy routing that classifies incoming recall queries into one of four strategy classes (exact_event_temporal, set_aggregation, count_query, default_mixed) and executes class-specific retrieval and reranking logic. Detection pipeline: - Step 1: vector + FTS search against control-plane recall_strategy_prototypes table, RRF merge, class-score aggregation with confidence thresholds - Step 2: LLM fallback (5s timeout) for low-confidence / ambiguous cases - Structural fallback to default_mixed when both steps are unresolved Execution model: - Per-strategy executors with entity-aware and answer-family-aware reranking - Fanout merge for the set_aggregation + count_query pair (60/40 budget split with dedup, backfill, and correct post-merge pagination) - All executors guard explicit-session paths on memory_type=="" to prevent returning session turns for insight-only queries Operational: - Router gated on TiDB backend with auto-embedding (not FTS-only) - Vector index bootstrap via EnsurePrototypeVectorIndex() at startup with duplicate-index error handling (rerunnable, no unconditional ALTER TABLE) - Prometheus counters for strategy/source/mode/fallback/fanout/duration - Bilingual prototype seed set (88 rows: en+zh across 4 strategy classes) Includes 21 focused unit tests covering step1/step2 threshold resolution, fanout pagination, memory_type guard regressions, and merge dedup semantics.
Preserve benchmark-provided message seq values in raw session storage so LoCoMo ingest keeps conversation-global ordering. Also tighten strategy router step-1 resolution by switching to weighted prototype similarity, tracking best class score, and falling back more often on weak prototype matches. Add focused tests for both paths.
Preserve the current Cat3-focused routing, retrieval, and answer-shape improvements as a rollbackable checkpoint before the next experiment round. This is not the final solution. It establishes a known-good baseline for subsequent benchmark-driven attempts, including explicit-session thread-chase and related Cat3 retrieval experiments.
Preserve the current Cat3 behavioral fixes as a rollbackable checkpoint before the next structural experiment. This round adds strategy-scoped explicit-session thread chase and related retrieval execution tuning, but it is not the final Cat3 solution. The next step may introduce a dedicated exact-entity lookup strategy if benchmark results still plateau.
Signed-off-by: Shenjun <[email protected]>
Preserve the current server-side Cat3 structural split after salvaging exact_entity_lookup to keep the raw question as the primary retrieval leg and use the shaped exact-entity query as additive support. This is a rollbackable benchmark checkpoint, not the final Cat3 solution.
👷 Deploy request for mem9 pending review.A Netlify team Owner will need to approve the deploy before you can run your build. Are you a team Owner? Visit the deploys page to approve it → Need more help? Learn more in the Netlify docs →
|
…ied-routed-recall-executor-proposal.md proposal Signed-off-by: Shenjun <[email protected]>
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
2 participants
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.
No description provided.