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: xerj code hybrid on a mixed-mapping corpus (#1146)
Vincenzo Lombardo5cc9c11f0abd
diff --git a/user-feedback/16-agent-field-reports/2026-10-06-xerj-code-hybrid-mixed-corpus.md b/user-feedback/16-agent-field-reports/2026-10-06-xerj-code-hybrid-mixed-corpus.mdnew file mode 100644index 000000000..86968fd77--- /dev/null+++ b/user-feedback/16-agent-field-reports/2026-10-06-xerj-code-hybrid-mixed-corpus.md@@ -0,0 +1,13 @@+# Reproducing and fixing `xerj code --mode hybrid` on a mixed-mapping corpus (2026-10-06)++**Agent:** Claude Code (Claude Opus 5.5) · **XERJ:** xerj v1.0.0-rc.83 (built from source) · **Platform:** macOS arm64++**Pointed at:** the #1146 repro, a 30-file Python git repo plus its corpus.json, on a clean `--data-dir` (61 records, 2 indices).++**Used it for:** reference coding (`xerj corpus add/index` → `xerj code`), to reproduce a bug and check the fix end to end.++**Verdict:** The corpus loop was easy to drive. `corpus add` and `corpus index` each printed the next command, and `xerj-done` gave an unambiguous end line. The issue's repro worked verbatim. The bug itself is the honest-claims kind: hybrid silently returned 1 hit where bm25 returned 20, and the `@mode` note claimed BM25 had covered both indices. Without the bm25 comparison I would have trusted it. The engine's per-index RRF made the fix clean: a single-leg `hybrid` gives scores on the same 1/(k+rank) scale, so the two responses merge by `_score` with no client-side fusion. One friction point: the installed binary was rc.74, which has no `xerj code` at all ("unknown argument: code"), so I could not get a same-version "before" without a second 7-minute release build. I replayed the old request with curl instead.++**Numbers:** `xerj code tinya "retry connection timeout handler" -k 20 --json`: bm25 20 hits; hybrid 1 hit before (old request via curl), 20 after. `xerj corpus index` wall=29.5s for 31 files. `cargo build --release -p xerj-server` 7m08s.++**Filed alongside:** PR #1177 (fix for #1146)