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.
Verified: the opener prompt works (2026-08-06)
The opener in OPENER_PROMPT.md was run verbatim as the instruction to a fresh
Claude Code agent, then independently verified (the verifier re-checked the
filesystem and XERJ, and compiled + tested the output — it did not trust the
executor's report).
That run used the wrapper scripts of the time (xc-index.sh, xc.py); they
were removed in #977, and the opener now names their built-in equivalents,
xerj corpus index and xerj code.
Scenario. A running XERJ at :9200, the xerj-code skill available, and one
reference library downloaded locally: sift, a real Rust crate whose API is
deliberately not new()/push()/top_k() but a two-phase, seal-gated count-min
sketch. The task: implement pub fn top3(stream: &[u64]) -> Vec<(u64, u32)>
returning the 3 heaviest keys — using sift's real API, which must be retrieved,
not guessed.
What the opener drove Claude Code to do:
- Index the reference —
xc-index.sh refs --fresh→ 3 records live (xc-refs-sift-src, …); confirmed byGET /_cat/indicesand_count → 3. - Retrieve before writing —
xc.py refs "sketch"returnedsift/src/lib.rs; the agent read the definition and used the real API. - Write correct code using
Sift::furnish(width, depth)/absorb/seal/crest— including theseal-before-read contract.
Independent verdict: WORKS (all four checks).
indexed✓ ·retrieved✓ ·used_real_api✓ ·compiles_and_correct✓The verifier built a fresh cargo crate depending on the real sift path, dropped in the generated
solution.rs, and rancargo test --release:test tests::matches_crate_readme_shape ... ok test tests::heaviest_three ... ok test result: ok. 2 passed; 0 failed; 0 ignoredasserting
top3([7,7,7,42,42,1,99]) == [(7,3),(42,2),(1,1)]andtop3([5,5,5,5,8,8,8,3,3,1]) == [(5,4),(8,3),(3,2)]— the exact ascending-key tiebreak sift documents.
Why this is strong evidence retrieval was actually used, not bypassed: sift's
defining feature is a seal-gated two-phase contract (read-before-seal panics,
absorb-after-seal panics, the crest is materialised at the seal) that
deliberately rejects the conventional new/push/top_k shape. A model working from
memory cannot invent this exact contract or the width-first/depth-second furnish
signature; code that respects it must have read the indexed source.
The generated program (verbatim):
use sift::Sift;
/// Return the 3 most frequent keys in `stream` as `(key, count)`, heaviest first.
pub fn top3(stream: &[u64]) -> Vec<(u64, u32)> {
let mut sketch = Sift::furnish(4096, 4); // width first, then depth
for &key in stream {
sketch.absorb(key);
}
sketch.seal(); // mandatory before any read
sketch.crest(3)
}