Semantic-only queries return 0 results despite active vector pipeline (query embedded, index populated) — v0.9.27

[verrecs]

Summary

On v0.9.27 (standalone/native mode), purely-semantic queries return 0 results from smart-search (compact mode) even though the whole vector pipeline is demonstrably active: the query is embedded at search time, the vector index is populated and persisted, and doctor reports the embedding provider green. Keyword (BM25) matches over memories work fine.

Environment

• agentmemory 0.9.27, standalone worker under systemd (user), native iii 0.11.2
• LLM + embeddings via Ollama's OpenAI-compatible API (local GPU):
  • OPENAI_BASE_URL=http://<lan-host>:11434, OPENAI_MODEL=qwen3:4b-instruct
  • EMBEDDING_PROVIDER=openai, OPENAI_EMBEDDING_MODEL=nomic-embed-text, OPENAI_EMBEDDING_DIMENSIONS=768
• Corpus: 52 imported Claude Code JSONL sessions → 23 780 observations, plus ~30 memories
• doctor: LLM provider ✓, Embedding provider ✓, consolidation ✓, compression ✓

What provably works

• Full index rebuild after purging stale index shards: 0 embed failed warnings, 39 vector shards persisted (mem:index:bm25:vectors:*)
• Query-time embedding: every smart-search call hits the embedding endpoint (verified via Ollama /api/ps keep-alive refresh)
• BM25: exact-keyword queries (e.g. a term present in a memory) return that memory

Bug

Semantic-only paraphrases (no shared tokens with the target text, e.g. querying in French "le noyau a assassiné une machine virtuelle faute de mémoire vive" against an OOM-incident memory that is semantically near-identical) return {"mode":"compact","results":[]} — via both the MCP tool and POST /agentmemory/smart-search.

Given HybridSearch.tripleStreamSearch() runs this.vector.search(queryEmbedding, limit*2) unconditionally once vector.size > 0, cosine top-N should always produce candidates, and RRF fusion should surface at least weak hits. It looks like either a post-fusion score threshold, or compact-mode formatting dropping observation-level (obsId) hits, hides all vector-leg results.

Two contributing traps hit on the way (worth docs/UX love)

1. Stale persisted index blocks the lazy rebuild forever: index shards persisted before a large import reload at boot → idx.size > 0 → the if (idx.size === 0) rebuildIndex() path never fires → imported observations are never indexed. AGENTMEMORY_DROP_STALE_INDEX=true only drops the vector side on dimension mismatch; BM25 shards keep the gate closed. We had to stop the worker and delete mem:index:* bins manually to force a full rebuild.
2. Dimensions env var naming: the OpenAI provider only reads OPENAI_EMBEDDING_DIMENSIONS; setting EMBEDDING_DIMENSIONS (a recognized var elsewhere) silently leads to Embedding dimension mismatch in openai.embedBatch[0]: expected 1536, got 768 and all rebuild batches being skipped.

Ask

• Where can vector-leg hits get filtered out between tripleStreamSearch fusion and compact-mode output?
• Would you accept a rebuild-index migrate step (or REST endpoint) to avoid the manual shard purge?

Happy to run any diagnostic build/flags — the setup is reproducible on demand.