🧪 test: Add Clean-Room Context Selection Replay

evals/__init__.py

@@ -0,0 +1 @@
+"""Isolated experiments, not production service modules."""

evals/context_selection/README.md

@@ -0,0 +1,211 @@
+# Context selection experiment
+
+This is an independently authored, opt-in replay harness, not a production
+search change. It tests whether selecting **answer-bearing evidence**, rather
+than always filling a top-k window, can reduce context without losing facts,
+exceptions, or opposing evidence.
+
+The motivation is GPT Researcher's context-filter write-up:
+https://docs.gptr.dev/docs/gpt-researcher/gptr/context-filter
+
+No upstream implementation or fixtures were copied. BM25 is the standard
+formula, with a deliberately simple NFKC/case-folded word tokenizer that retains
+negation and identifiers. We previously inspected upstream source, so this is
+not a claim of a formal legal clean-room process. The provider protocol is
+interoperability, not a dependency on GPT Researcher or its classes.
+
+## Boundary and ownership
+
+```text
+fixed, already-authorized passage pool + query
+  -> score once (BM25 / embeddings / Jev)
+  -> selection ablations using those same scores
+  -> original passage IDs and sources + evidence-retention metrics
+```
+
+The harness imports neither `app.config` nor application storage. Production
+routes, schemas, chunking, caches, authorization and ingestion are untouched.
+There is no server endpoint, database probe, scrape, writer LLM or automatic OCR.
+The operator owns corpus authorization and permission to send its content to a
+provider. Never use private production documents without that authorization.
+
+A candidate is one complete existing chunk. This experiment does not re-chunk
+or summarize it. Keep candidate generation constant across strategies; later
+chunk-size experiments need separately fingerprinted corpora. A selector cannot
+recover evidence omitted by the retriever.
+
+Selection owns stable tie ordering, thresholds and output budgets. The HTTP
+adapter owns response validation, shared concurrency and cancellation. Evaluation
+labels are used only after selection and are never included in provider payloads.
+
+## Strategies
+
+| Strategy | Behavior |
+| --- | --- |
+| `all` | Original pool order, subject to the character budget, not top-k |
+| `input-topk` | Original retrieval order, top-k and character budget |
+| `bm25-topk` | Lexical scoring without a relevance threshold |
+| `bm25-filter` | Lexical scoring, relative cutoff against the best score |
+| `embeddings-topk` | Cosine ranking without a threshold |
+| `embeddings-filter` | Cosine ranking with an absolute threshold |
+| `jev-topk` | Usefulness ranking without a threshold |
+| `jev-filter` | Usefulness ranking with an absolute threshold |
+
+`max_results` is a ceiling, not a requirement to fill slots. Relative BM25
+selection returns empty when no term matches. Successful Jev rejection returns
+empty, never the first passages as a disguised fallback. A provider failure is
+an **error**, not a successful lexical run attributed to the provider. Provider
+errors make the command exit 1; skipped runs are explicitly counted.
+
+Thresholds are experimental settings, not demonstrated calibration. Jev scores
+are expected rubric levels in `[0, 3]`, not answer-correctness probabilities.
+BM25 and cosine scores are not on that scale. There is no evidence that one
+threshold fits every query type, language or model.
+
+`max_chars` counts Python Unicode characters in complete selected passage
+bodies. It is **not** a model token budget, UTF-8 byte budget or final prompt
+budget. Citation wrappers also cost tokens. Oversized passages are skipped,
+not truncated into apparently complete evidence. Optional exact-text
+deduplication applies to scored strategies; `all` and `input-topk` remain
+unmodified controls. It does not merge overlapping or paraphrased passages.
+
+## Run offline
+
+Python 3.12 is the repository's CI version. The local CLI uses only the standard
+library. HTTP checks need `httpx` from `test_requirements.txt`.
+
+```sh
+mkdir -p .venv
+python3 -m evals.context_selection.replay \
+  --max-results 2 --output .venv/context-offline.json
+```
+
+The default includes BM25 and explicitly **skips** embeddings and Jev. Even if
+keys exist, no client is constructed without `--allow-network`. There is no
+`.env` loading. Missing live configuration is also recorded as skipped, never
+silently replaced with simulated model scores.
+
+The committed 10 cases and 82 passages are synthetic boundary fixtures:
+refund exceptions, exact issue identifiers, negation, multi-facet answers,
+conflicting dated sources, Unicode, duplicate passages, evidence after the
+first 50 candidates, irrelevant pools, and empty pools. They establish behavior
+and expose lexical blind spots. They are **not** a representative benchmark or
+an independent estimate of answer quality.
+
+## Live replay, once credentials are provided
+
+Set these through the environment or secret tooling, not tracked files or PR
+comments:
+
+- `CONTEXT_EMBEDDING_URL`: full OpenAI-compatible embeddings endpoint URL.
+- `CONTEXT_EMBEDDING_MODEL`: explicit model name.
+- `CONTEXT_EMBEDDING_KEY`: provider key.
+- `TYPESAFE_API_KEY`: Jev provider key.
+- `CONTEXT_JEV_MODEL`: defaults to `jev-latest`; pin a model when possible.
+- `CONTEXT_JEV_URL`: optional, defaults to the System One endpoint.
+
+The URLs are operator-controlled endpoints, not untrusted user URL inputs.
+TLS verification remains enabled and redirects are not followed. Secrets and
+provider response bodies are not written to reports or printed on failures.
+Queries and passages go to the configured provider only on explicitly enabled
+runs. Score outputs include models, indices, call counts, reported input tokens
+(or null if absent), and measured scoring time. Failed batches retain attempted
+request counts and the subtotal of input tokens reported before failure, but
+mark total usage unknown. No pricing is invented.
+
+```sh
+python3 -m evals.context_selection.replay --allow-network \
+  --max-results 2 --output .venv/context-live.json
+
+# Sweep selection settings without paying for the same scores again.
+python3 -m evals.context_selection.replay \
+  --reuse-scores .venv/context-live.json --jev-min 2.0 \
+  --max-results 4 --output .venv/context-sweep.json
+```
+
+Corpus bytes and schema version must match for reuse. Scoring happens once per
+case/provider; top-k and filtering reuse that result. Costs/calls must be read
+from `score_runs`, not summed across selection rows. A saved-score replay
+records zero new provider calls. Do not compare its inherited scoring latency
+to a new network run.
+
+Jev uses a fixed number of worker tasks and a semaphore shared across calls on
+that adapter. A batch deadline includes queue wait. Failure or cancellation
+cancels and drains sibling tasks. We intentionally make no automatic retries;
+rerun a failed experiment explicitly. The per-case candidate/character bounds
+are enforced before any provider call. The default embedding batch contains
+query plus candidate text; provider-specific batch limits may be lower and
+should be reflected in the flags. Do not silently drop excess candidates.
+
+## Corpus and metrics
+
+`--corpus` accepts a JSON array of cases. Passage order is the original
+retrieval order. IDs must be unique within a case; case IDs are globally unique.
+Each required fact has an evidence label; any passage may support several.
+Every required label must be present somewhere in the candidate pool, so
+selection recall is not confused with retrieval recall. A `required: []` case
+means no candidate contains useful answer evidence, not that the real-world
+question has no answer.
+
+```json
+[
+  {
+    "id": "retention-example",
+    "query": "When can a record be deleted?",
+    "required": ["duration", "legal-hold"],
+    "passages": [
+      {"id": "p1", "source": "policy:page1", "text": "Keep records for seven years.", "supports": ["duration"]},
+      {"id": "p2", "source": "policy:page2", "text": "A legal hold prevents deletion.", "supports": ["legal-hold"]}
+    ]
+  }
+]
+```
+
+Reports retain IDs and source handles, not query or passage bodies. They contain:
+
+- evidence precision: fraction of selected passages carrying annotated evidence;
+- evidence recall: fraction of required labels retained;
+- missing-evidence labels, including exceptions and contradicting sources;
+- correct abstentions for pools with no annotated evidence;
+- selected characters and passage counts;
+- separate successful, skipped and error counts.
+
+Precision is null for empty selections, recall is null for no-evidence cases.
+Macro summaries report their denominators and exclude null values, so inspect
+missing-evidence and abstention counts alongside them. High precision alone can mean useful facts were lost.
+This is label-based selection evaluation, not generated-answer faithfulness.
+
+## Verify this slice
+
+```sh
+# The harness needs no application fixtures, database, or embedding initialization.
+python -m pytest --noconftest tests/test_context_selection.py
+python -m black --check evals/context_selection tests/test_context_selection.py
+python -m compileall -q evals/context_selection tests/test_context_selection.py
+```
+
+The test file is also discovered by the repository's normal CI unit-test run.
+MockTransport emulates external APIs; failure/cancellation tests exercise the
+actual adapter tasks and cleanup, not a fake implementation of selection.
+
+## Acceptance before production integration
+
+1. Replay representative, authorized LibreChat web/file candidate pools with
+   human-reviewed evidence labels. Keep a held-out set for threshold selection.
+2. Compare identical candidates, budgets, and writer models if answer generation
+   is added. Track identifier queries, broad questions, caveats, conflicts,
+   multilingual content and answer evidence late in the pool separately.
+3. Require retained evidence and citation provenance, not token reduction alone.
+   Record provider latency, inference usage and downstream prompt-token counts
+   with the actual tokenizer before claiming speed or cost wins.
+4. Exercise the real adapters with provider credentials. MockTransport tests
+   prove our expected HTTP contract, not current provider behavior or calibration.
+5. Only then add an opt-in production selector. Existing search is the rollback
+   path. Auth/tenant/entity scope must be checked before inference egress, and
+   filtering must never substitute for whole-file content inspection.
+6. Update the consuming agents contract to distinguish an honest empty evidence
+   set from service failure. Its current RAG reranker treats empty results as a
+   bad response and falls back; that must not be reused unchanged for filtering.
+
+No model quality, calibration, answer-quality, real-provider latency, downstream
+cost reduction, or production no-regression claim follows from the offline run.

evals/context_selection/__init__.py

@@ -0,0 +1 @@
+"""Clean-room context-selection experiment."""