retrieval-fairness 0.1.1

Exposure audit for vector search / RAG: coverage, dark matter, Gini, hub capture, CI gate

retrieval-fairness

Exposure audit for vector search / RAG. Shows what share of your vector corpus is actually reachable by queries and what share is never retrieved (dark matter — the antihub inventory of your index); measures exposure concentration (Gini), hub capture, and regression diff when you change the embedder or chunking. Think of it as a health/coverage report for your index, not your pipeline.

Status: early development. This is packaging novelty (the Gini / retrievability metrics are honestly borrowed from IR-fairness / T-Retrievability research), not a from-scratch invention.

Why your index has dark matter (it's not a bug in your code)

Hubs and never-retrieved chunks are a structural property of high-dimensional nearest-neighbor search, not a symptom of a bad embedder. Radovanović, Nanopoulos & Ivanović (JMLR 2010, “Hubs in Space”) showed that as dimensionality grows, the distribution of k-occurrences becomes strongly right-skewed: a few points (hubs) appear in a disproportionate share of neighbor lists, while others (antihubs) appear in almost none. Modern ANN indexes (HNSW) amplify the effect — hub nodes are the highway entry points that make the graph navigable. Left unmeasured, this silently collapses the effective size of your corpus. This tool makes it measurable — and gateable in CI.

Installation

pip install retrieval-fairness

Optional store adapters are extras (installed only if you use them):

pip install 'retrieval-fairness[faiss]'       # FAISS
pip install 'retrieval-fairness[pgvector]'    # pgvector (PostgreSQL)
pip install 'retrieval-fairness[qdrant]'      # Qdrant
pip install 'retrieval-fairness[models]'     # sentence-transformers embedder
pip install 'retrieval-fairness[fastembed]'   # fastembed (BGE) embedder

From source (development):

pip install -e '.[dev]'

Quick start

retrieval-fairness demo --top-k 5          # demo on a synthetic corpus
retrieval-fairness demo-diff --top-k 5     # regression diff for an embedder migration

Usage

Run against real queries

# corpus.jsonl: {"id": "...", "text": "...", "vector": [...]}
# queries.jsonl: {"id": "...", "vector": [...]}
retrieval-fairness probe --corpus corpus.jsonl --queries queries.jsonl \
    --top-k 10 --json report.json --html dashboard.html

No query logs: antihub self-query audit (synthetic queries)

For each chunk, generate the query that should retrieve it (its own top TF-IDF terms) and check whether it actually surfaces. A chunk that cannot be found even by a query aimed at it is invisible from any reasonable query direction — dark matter from day one:

retrieval-fairness synth --corpus corpus.jsonl --top-k 10 --html dashboard.html

Regression diff (embedder/chunking change)

retrieval-fairness diff --baseline before.json --candidate after.json
# For an intentional rechunking migration (different corpus fingerprint):
retrieval-fairness diff --baseline before.json --candidate after.json \
    --corpus-policy allow-change

Baselines use schema v2 and contain ordered query IDs plus deterministic corpus and workload SHA-256 fingerprints. On load, reports are rebuilt from raw hits and frequencies; saved metric values are never trusted as source data. Legacy 0.1.0 baselines remain readable, using positional query alignment only when explicitly allowed by strict overlap gates.

CI gate

retrieval-fairness gate --baseline v1.json --candidate new.json --strict \
    --max-coverage-drop 0.05 --max-dark-matter-rise 0.05
# exit 1 in strict mode if coverage dropped > 5 pp -> CI blocks the deploy

Cross-check dark matter against qrels ("lost gold")

If you have relevance judgments (qrels), cross-check which dark-matter chunks are actually relevant — the corpus contains material the retriever never surfaces. No competitor exposure tool ships this:

retrieval-fairness qrels --probe report.json --qrels qrels.json \
    --min-relevance-grade 1 --json qrels_report.json
# --queries is only required for a legacy schema-v1 probe

A qrels pair is relevant only when grade >= --min-relevance-grade (default 1), so zero and negative judgments are ignored. The output reports micro recall@k over all relevant query/document pairs and macro recall@k over queries that have at least one relevant in-corpus document. recall_at_k is retained as a compatibility alias for micro recall for the 0.1.x cycle.

Metrics

Metric	What it shows
Coverage %	share of the corpus retrieved at least once
Of reachable ceiling %	coverage as a share of what the workload can physically reach (n_queries × top_k); distinguishes a bad retriever from a small workload
Dark matter %	share NEVER retrieved
Gini	exposure concentration (0 = uniform, 1 = all in one)
Hub capture top5/10	share of exposure captured by top-N hubs
Lorenz curve	inequality visualization
Per-query overlap	result stability across a migration
Lost gold / micro & macro Recall@k	positive-relevance dark-matter chunks and qrels retrieval quality

How it works

• retrieval_fairness/types.py — the VectorStore contract (Protocol). Any store is bridged to it via an adapter (InMemory, FAISS, pgvector, Qdrant today; Pinecone/Weaviate on the roadmap).
• adapters/inmemory.py — InMemoryVectorStore (cosine, numpy) for dev/tests/demos; adapters/{faiss,pgvector,qdrant}.py — store adapters.
• metrics.py — coverage, gini, lorenz, hub_capture, FairnessReport.
• probe.py — run a workload → retrieval frequency → report.
• diff.py — regression diff between two runs.
• gate.py — CI gate with configurable rules.
• synth.py — antihub self-query audit (synthetic queries from the corpus).
• qrels.py — dark-matter vs qrels cross-check ("lost gold").
• dashboard.py — self-contained HTML report (Lorenz, histogram, PCA map).
• embedders.py — Embedder contract (TF-IDF / sentence-transformers / fastembed).

Real-scale case study (BEIR NQ, ~50% dark matter, lexical→dense regression diff): docs/case_study_nq.md. Store adapters: docs/adapters.md. Comparison with related work: docs/comparison.md.

Tests

pytest tests/ -q

License

MIT