generated from hack-ink/vibe-mono
-
Notifications
You must be signed in to change notification settings - Fork 1
Open
Labels
area:serviceRetrieval logic, ranking, and request orchestration.Retrieval logic, ranking, and request orchestration.area:storagePostgres schema, SQL queries, and storage correctness.Postgres schema, SQL queries, and storage correctness.area:workerBackground workers, outbox processing, and indexing pipelines.Background workers, outbox processing, and indexing pipelines.kind:epicUmbrella issue that tracks multiple deliverables.Umbrella issue that tracks multiple deliverables.theme:evaluationQuality measurement, gold sets, regressions, and metrics.Quality measurement, gold sets, regressions, and metrics.
Description
Context
We are redesigning the memory retrieval architecture to match the strongest known patterns from qmd and claude-mem, while keeping ELF invariants (evidence binding, deterministic writes, and rebuildable indexing).
Goal
Deliver a best-in-class retrieval pipeline focused on recall, precision, ranking stability, and controllable context growth.
Scope (tracked by sub-issues)
- Implement query expansion with multi-query hybrid retrieval and RRF fusion #12 Implement query expansion with multi-query hybrid retrieval and RRF fusion
- Add chunked embeddings with overlap and chunk stitching #13 Add chunked embeddings with overlap and chunk stitching
- Add context metadata and context-aware boosting in retrieval #14 Add context metadata and context-aware boosting in retrieval
- Normalize rerank scores and apply position-aware blending #15 Normalize rerank scores and apply position-aware blending
- Add optional note-level diversity control (MMR / redundancy filter) #28 Add optional note-level diversity control (MMR / redundancy filter)
- Add optional deterministic ranking signals (lexical overlap, hit reinforcement, decay) #29 Add optional deterministic ranking signals (lexical overlap, hit reinforcement, decay)
- Introduce progressive disclosure search endpoints (index, timeline, details) #16 Introduce progressive disclosure search endpoints (index, timeline, details)
- Add structured memory fields with field-level embeddings #17 Add structured memory fields with field-level embeddings
- Graph-lite entities and relations as structured memory fields (no graph DB) #27 Graph-lite entities and relations as structured memory fields (no graph DB)
- Add LLM cache for query expansion and reranking #18 Add LLM cache for query expansion and reranking
- Add search explainability outputs #19 Add search explainability outputs
- Build a retrieval evaluation harness and regression gates #20 Build a retrieval evaluation harness and regression gates
Backlog (explicitly out of scope for this epic)
- Add a read-only web viewer for sessions and traces #26 Add a read-only web viewer for sessions and traces (deferred)
Acceptance criteria
- All in-scope sub-issues are delivered.
- Each change that affects ranking is evaluated with the retrieval harness (same dataset, before/after comparison).
References
Reactions are currently unavailable
Metadata
Metadata
Assignees
Labels
area:serviceRetrieval logic, ranking, and request orchestration.Retrieval logic, ranking, and request orchestration.area:storagePostgres schema, SQL queries, and storage correctness.Postgres schema, SQL queries, and storage correctness.area:workerBackground workers, outbox processing, and indexing pipelines.Background workers, outbox processing, and indexing pipelines.kind:epicUmbrella issue that tracks multiple deliverables.Umbrella issue that tracks multiple deliverables.theme:evaluationQuality measurement, gold sets, regressions, and metrics.Quality measurement, gold sets, regressions, and metrics.