AI search engine and MCP server for bots and agents. One Rust binary indexes code, docs, logs, JSONL/CSV and PDFs for BM25 full-text, vector and hybrid search, code search without grep, agent memory and RAG. Elasticsearch-compatible API; Corpus Hub serves hundreds of reference memory corpora.
Rust agent-memoryai-agentsbm25claude-codeclicode-searchcursordatabaseelasticsearchfull-text-searchhybrid-searchjsonllocal-firstmcpmcp-servermodel-context-protocolragrustsearch-enginevector-search
field report: diagnosing slow semantic/hybrid search on BEIR FiQA (#1091)
Vincenzo Lombardo2c1941b649a6
diff --git a/user-feedback/16-agent-field-reports/2026-10-07-fiqa-semantic-exact-scan.md b/user-feedback/16-agent-field-reports/2026-10-07-fiqa-semantic-exact-scan.mdnew file mode 100644index 000000000..4dc90b243--- /dev/null+++ b/user-feedback/16-agent-field-reports/2026-10-07-fiqa-semantic-exact-scan.md@@ -0,0 +1,13 @@+# Diagnosing and fixing slow semantic/hybrid search on BEIR FiQA (2026-10-07)++**Agent:** Claude Code (Claude Opus 5.5) · **XERJ:** xerj v1.0.0-rc.87 (built from source), rebased onto rc.88 · **Platform:** macOS arm64 (Apple M5, 24 GB)++**Pointed at:** BEIR FiQA (57,638 docs, `semantic_text` body) loaded with `benchmarks/beir-hybrid/load.py` on a throwaway node.++**Used it for:** vector and hybrid search: reproducing #1091 (13-26 s hybrid queries), then A/B-testing the fix.++**Verdict:** The diagnosis was easier than the issue suggested. It reproduced with the default lexical embedder, so the neural model was a red herring, and `XERJ_TRACE_SEMANTIC_PHASES` plus the `segment_hydration_cache` block in `_nodes/stats` showed where the time went: per-segment stored-document decodes, a full hydration budget, and the memory-watermark drain wiping the caches. The real cost is opacity. Nothing in a normal response says "this field cannot use HNSW because some documents are passage-chunked", so a user sees a slow query with no reason; the guard's exit is only visible by instrumenting it. Two traps cost me time. A flush splits across memtable shards, so small segments are LZ4 and the typed projection does not apply to them. And `load.py` stopped at the first 429 from the memory breaker, so my first A/B silently compared a 36k-doc index against a 57k one. The storage layer already had exactly the primitives the fix needed (typed kNN projection, row-selective hydration), unused by the engine.++**Numbers:** FiQA, 50 queries per arm, matched release builds of `e4c6c027` and the fix: semantic p50 3,865 → 261 ms, hybrid p50 5,772 → 289 ms, memory-watermark crossings 25 → 0, ranked ids identical 50/50 in both arms. ES-YAML on the final head: 1385 passed, 0 failed. `cargo build --release -p xerj-server` 4-8 min.++**Filed alongside:** PR #1217 (fix for #1091)