Baseline-first ordering 10: an opt-in nonlearned search helper

한국어 · Full results

Keep the baseline's first candidate and interleave identifier-split suggestions afterward. On 2,948 documentation-proxy questions this preserves Recall@1 and improves Recall@10/mean target rank. This is a nonlearned baseline, not a learned-model success. Defaults remain unchanged.

Try it

No model download or Python is needed for this original public fixture:

printf '%s\n' '{"snapshot_id":"demo","query":"read config","documents":[{"id":"a","text":"read"},{"id":"b","text":"config"},{"id":"c","text":"func ReadConfig() {}"}]}' \
  | go run ./cmd/riido-hints --identifier-hints

Baseline order a,b,c becomes a,c,b. Candidate a stays first; splitting ReadConfig into read config introduces c second. Output contains IDs, ranks and original BM25 scores, not document text. Original scores need not decrease in the interleaved order. Status is unverified, policy identifier_bm25_baseline_first.

Omit the flag for baseline behavior. Combining it with external hints or --model is rejected. If normalized input exceeds existing limits or contains no searchable query words, retain the baseline with policy bm25_identifier_input_out_of_scope. Callers must supply only authorized documents; the tool does not crawl repositories or decide permissions. It currently emits all candidates, so output-token costs for large catalogs still need separate consideration.

Boundaries

The Go API is hintsearch.InterleaveBaselineFirst(baseline, hints). Partial hints are allowed; all baseline candidates remain without duplicates. A candidate at baseline rank r appears by min(n,2r−1), so baseline rank 3 appears by rank 5. This bounds candidate checks under accurate equal-cost verification, not wall time, tokens, multi-evidence tasks or hint construction. Preserving the first candidate does not prevent regressions later in the ranking.

Corpus results

Same three pinned repositories, 3,009 functions and 2,948 distinct descriptions as 07/09. Exclusions, candidate hash ordering and known targets were unchanged. Implementation was committed before first corpus scoring; replay reproduced the complete JSON byte-for-byte. This previously observed documentation proxy is not a fresh independent user-query test.

Policy	Recall@1	Recall@10	MRR	Mean first-target rank
Raw BM25	41.79%	74.83%	0.5336	48.99
Auxiliary-first interleave	39.21%	81.51%	0.5479	26.71
Baseline-first interleave	41.79%	81.38%	0.5601	26.70

Baseline-first improves 733 questions, worsens 674 and leaves 1,541 unchanged. Average gain is not universal gain. Per-repository metrics are in JSON. Learned models must beat this nonlearned control while accounting for additional cost. No experiment-09 failure was promoted and no HF weights were replaced.

Validation and cost

Exhaustive tests for 1–7 candidates cover every hint permutation/partial length, first/full-candidate preservation and the 2r−1 bound. Existing auxiliary-first behavior remains tested. Re-running the experiment-09 trainer through the updated shared implementation reproduced its complete report. Actual CLI execution and tests cover the example, conflicting inputs and out-of-scope fallback.

Apple M4 Pro/Go1.27.1 CPU: the whole comparison took 3.02s wall/2.67s user CPU, peak RSS 94,683,136 bytes. This includes source validation and two complete search preparations for comparing policies; it is not single-CLI/service latency. The option builds an additional index and search, so end-to-end time savings are not guaranteed. No training, GPU or LLM call occurs in this comparison.

# Uses the pinned corpus prepared for experiment 07.
go run ./cmd/riido-interleaveprobe
go test -race ./pkg/hintsearch ./cmd/riido-hints ./internal/retrievalbench

Next training must account for inspection gain and auxiliary-search cost relative to this control. Actual user-task success, end-to-end time/token savings, independent questions and reserved CoSQA evaluation remain outstanding.