Skip to content

armgit / xerj

Public Imported from GitHub
0

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.

docs(field-report): agent reference-coding report 2026-10-04

Alessandro Costab5c43ae5251d
diff --git a/user-feedback/16-agent-field-reports/2026-10-04-node-oss-reference-coding.md b/user-feedback/16-agent-field-reports/2026-10-04-node-oss-reference-coding.md
new file mode 100644
index 000000000..70ce0dad1
--- /dev/null
+++ b/user-feedback/16-agent-field-reports/2026-10-04-node-oss-reference-coding.md
@@ -0,0 +1,13 @@
+# Reference-coding over a Node app's OSS deps on a constrained box (2026-10-04)
+
+**Agent:** Claude Code (Opus 4.8), on behalf of a human maintainer · **XERJ:** v1.0.0-rc.80 (x86_64 musl) · **Platform:** Linux x86_64
+
+**Pointed at:** Public OSS deps of a Node/WhatsApp codebase — Baileys (~160 code files) and ioredis — on a RAM-tight shared edge box (7.6 GB, no GPU), the node capped via a systemd `--user` cgroup (CPUQuota, MemoryMax, MemorySwapMax=0).
+
+**Used it for:** autoindex + lexical BM25 query, and the `corpus` / `xerj code` reference-coding path.
+
+**Verdict:** `autoindex` was the strong point — reliable and fast, and right: `baileys` indexed in 84.6s and `?q=sendMessage` ranked `messages-send.ts` first (the file that defines it). The `corpus index` path was the opposite on this host: it never finished the finalize/switch (a ~490-file corpus ran >2h at CPUQuota=300% and was killed by my 7200s service timeout), and a corpus indexed against one `--url` then queried on another failed with `state directory cannot become generation`, leaving 0 live indices plus orphaned `xc-*` generations. I could not cleanly isolate the hang from my own constraints (cgroup caps, repeated kills, RAM pressure), so I am not filing it as an issue; the stale-generation half looks like the open #1136. Net: on a constrained box I would reach for `autoindex` + `xerj_search` again and avoid the `corpus` lifecycle until generation/state handling settles.
+
+**Numbers:** `xerj autoindex .../baileys --prefix baileys` -> ok=true, wall=84.6s, files=163, records=3074, code_files_indexed=143. `xerj corpus index ioredis` -> killed at the 7200s service timeout, never switched. autoindex idle RSS ~36 MB (lexical), ~148 MB with the neural model loaded.
+
+**Filed alongside:** no cleanly-reproducible defect beyond the open #1136 (corpus generation leaks); this report is the record.