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): local /_decide email-classification prototype session

Vincenzo Lombardocb7cd2e9c329
diff --git a/user-feedback/16-agent-field-reports/2026-09-23-systemone-decide-email-classify-proto.md b/user-feedback/16-agent-field-reports/2026-09-23-systemone-decide-email-classify-proto.md
new file mode 100644
index 000000000..e00881136
--- /dev/null
+++ b/user-feedback/16-agent-field-reports/2026-09-23-systemone-decide-email-classify-proto.md
@@ -0,0 +1,13 @@
+# Prototyping local email classification via `/_decide` (2026-09-23)
+
+**Agent:** Claude Code (Sonnet 5) · **XERJ:** xerj v1.0.0-rc.76 · **Platform:** macOS arm64 (Darwin 27.0.0)
+
+**Pointed at:** a self-generated synthetic corpus, 260 short emails, 4 labels (invoice/newsletter/personal/needs-reply) — not a real mailbox.
+
+**Used it for:** `/_decide` (the local, no-egress k-NN "System One" vote over a labelled-history index) as a classification prototype for discussions/1012.
+
+**Verdict:** `/_decide` was trivial to drive from a throwaway Python script — no client library, one POST per item, sane response shape (label/confidence/abstain/neighbours). It worked exactly as documented in `systemone_api.rs`'s module comment on first try, no surprises, no doc gap found. The one thing worth a maintainer's attention isn't a bug: the confidence score is not a reliable abstain signal on inputs that don't lexically resemble the labelled history — it was wrong *and* confident (0.84–1.0) on 3/8 deliberately hard test items, same miscalibration shape #1012 already flags for the hosted path. Would use `/_decide` again for a quick local classifier; would not ship its confidence as a threshold without tuning against real data first.
+
+**Numbers:** `python3 run.py` against a local `xerj -d ./data --insecure --port 9420` → easy tier 60/60 = 1.000 acc, hard tier 5/8 = 0.625 acc, 2.6ms/item and 1.7ms/item respectively. Full run: `benchmarks/systemone-classify-email/results/run.json`.
+
+**Filed alongside:** [PR #1026](https://github.com/xerj-org/xerj/pull/1026) (the benchmark) and a comment on [discussions/1012](https://github.com/xerj-org/xerj/discussions/1012#discussioncomment-18562551).