VecAdvisor

Cost-based CLI advisor for filtered vector search in PostgreSQL and pgvector.

Read the story: Filtered vector search in pgvector can silently lose recall

VecAdvisor helps answer a question PostgreSQL does not currently cost well:

SELECT id
FROM documents
WHERE tenant_id = 42
ORDER BY embedding <-> $query_vector
LIMIT 10;

When a vector index is searched first and SQL filters are applied after the ANN candidate set, selective filters can silently lose recall or return fewer than k rows. Sometimes a filter-first exact scan is faster and exact. Sometimes pgvector iterative scans are the right answer. Sometimes a partial HNSW index or partitioning is the durable fix.

Asciinema cast fallback

This project models that choice with:

• PostgreSQL catalog and statistics introspection.
• Global selectivity estimation and PostgreSQL plan-row cross-checks.
• Local selectivity probes over the query vector neighborhood.
• A calibrated cost model for exact, post-filter ANN, iterative ANN, partial index, and partition-pruned strategies.
• Reproducible synthetic benchmark, calibration, validation, crossover, and proof reports.

Status: alpha. The Python CLI is useful today for experimentation and design validation. It does not modify PostgreSQL's planner.

VecAdvisor is an independent third-party tool. It works with PostgreSQL and pgvector, but it is not affiliated with the official pgvector project.

• Design
• Predicate support
• Workload collector
• Native kernels
• Rust/pgrx extension
• Benchmark artifacts
• SIFT1M scale benchmark recipe
• Release checklist

Global selectivity is the fraction of the whole table matching a filter. Local selectivity is the fraction of near-neighbor candidates matching the filter for a query vector.

For filtered ANN, local selectivity is the important signal. A tenant may be globally rare but locally dense near its own documents. Or it may be locally sparse for a query outside that tenant's embedding cluster. A naive k / s model misses this. VecAdvisor probes a bounded unfiltered top-m neighborhood and costs durable recommendations against p10 local selectivity across representative query vectors.

Full write-up of these results: Filtered vector search in pgvector can silently lose recall

The lead artifacts measure actual PostgreSQL/pgvector behavior, not a self-contained simulator. The scale artifact uses 1,000,000 real SIFT vectors and is the strongest first read:

With global filter selectivity fixed at 5%, the anti-correlated SIFT1M run has zero passing rows in every query's exact top-40 neighborhood. Fixed-frontier postfilter reaches only 0.3438 recall@k and returns full k for 25% of queries. Iterative HNSW recovers 1.0000 recall@k and full k, a 65.6 percentage-point recall improvement over fixed postfilter.

The companion Pareto chart is still useful for throughput/quality tradeoffs: SIFT1M anti-correlated Pareto.

See docs/benchmarks/sift1m-anticorrelated-pgvector-benchmark.md for the full SIFT1M recall-collapse artifact. A companion projection-tail SIFT1M run is also committed; it reaches 0.1750 postfilter recall@k and 0.9875 iterative recall@k: docs/benchmarks/sift1m-pgvector-benchmark.md.

A smaller Docker-reproducible benchmark keeps the SQL strategy mechanics easy to inspect:

Using the bundled pgvector/pgvector:pg17 container, fixed-size HNSW post-filtering reached only 0.2125 recall@k and returned full k for 0% of queries on a filtered workload. Iterative, partial-index, and partition-pruned HNSW recovered result quality. This is the failure mode VecAdvisor is designed to catch before it becomes a production surprise.

See docs/benchmarks/real-pgvector-benchmark.md for the SQL strategy details, hardware notes, and reproduction commands.

On the measured PostgreSQL crossover sweep, VecAdvisor recommended a recall-target-meeting plan in every failed-postfilter bin (9/9). It selected the exact latency-optimal plan in 4/9 bins; in the remaining bins it chose a recall-safe, slightly slower plan. Safety is the MVP1 claim; exact latency-optimal ranking is a calibration roadmap item: real-pgvector-crossover.

The repository also includes deterministic synthetic validation under docs/benchmarks/:

That sweep varies filter selectivity (0.01, 0.05, 0.1, 0.3) and filter/vector correlation (-0.6, 0, 0.6) across 12 points. It is a unit-level model-vs-simulation agreement check: the cost-model winner matched the simulated strategy-semantics winner in all bins, and default post-filter ANN missed recall or returns-k targets in all bins. Treat it as model validation, not proof of production pgvector latency.

Reproduce the synthetic artifact:

vecadvisor calibrate \
  --dataset synthetic \
  --rows 2048 \
  --dim 32 \
  --queries 24 \
  --clusters 8 \
  --filter-selectivity 0.1 \
  --correlation 0.5 \
  --limit 10 \
  --ef-sweep 10,20,40,80,160 \
  --seed 410 \
  --dataset-id synthetic-readme \
  --hardware-id github-readme-synthetic \
  --out docs/benchmarks/synthetic-calibration.json

vecadvisor benchmark-sweep \
  --dataset synthetic \
  --rows 2048 \
  --dim 32 \
  --queries 24 \
  --clusters 8 \
  --filter-selectivities 0.01,0.05,0.1,0.3 \
  --correlations -0.6,0,0.6 \
  --limit 10 \
  --ef-search 40 \
  --max-scan-tuples 1000 \
  --probe-rows 200 \
  --calibration docs/benchmarks/synthetic-calibration.json \
  --seed 411 \
  --out docs/benchmarks/synthetic-sweep.json

vecadvisor proof docs/benchmarks/synthetic-sweep.json \
  --out docs/benchmarks/synthetic-proof.json

vecadvisor plot-crossover docs/benchmarks/synthetic-sweep.json \
  --out docs/assets/synthetic-crossover.svg \
  --title "VecAdvisor synthetic crossover"

Install from PyPI:

python -m pip install vecadvisor

For local development:

python -m pip install -e ".[dev]"

The current alpha release is also available as wheel and source artifacts on GitHub:

python -m pip install https://github.com/vecadvisor/vecadvisor/releases/download/v0.1.0a4/vecadvisor-0.1.0a4-py3-none-any.whl

The CLI entry point is:

vecadvisor --help

From a checkout:

python -m pip install -e ".[dev]"
docker compose -f docker/docker-compose.yml up -d
docker compose -f docker/docker-compose.yml exec -T postgres \
  psql -U postgres -d vecadvisor < examples/demo.sql

vecadvisor explain \
  --dsn postgresql://postgres:postgres@localhost:5432/vecadvisor \
  --table public.documents \
  --vector embedding \
  --query "tenant_id = 1" \
  --q-vector examples/query-vector.json \
  --probe-rows 16 \
  --format text

vecadvisor recommend \
  --dsn postgresql://postgres:postgres@localhost:5432/vecadvisor \
  --table public.documents \
  --vector embedding \
  --query "tenant_id = 1" \
  --q-vectors examples/query-vectors.json \
  --probe-rows 16 \
  --max-query-vectors 3 \
  --format text

No representative query vectors yet? Use the table-sampled fallback for a 30-second first look. VecAdvisor marks this path as low confidence:

vecadvisor recommend \
  --dsn postgresql://postgres:postgres@localhost:5432/vecadvisor \
  --table public.documents \
  --vector embedding \
  --query "tenant_id = 1" \
  --allow-table-sample-vectors \
  --format text

Captured outputs:

• examples/explain-output.txt
• examples/recommend-output.txt

Start the bundled pgvector database:

docker compose -f docker/docker-compose.yml up -d

Load the example table:

psql "postgresql://postgres:postgres@localhost:5432/vecadvisor" -f examples/demo.sql

If you do not have psql locally, use Docker:

docker compose -f docker/docker-compose.yml exec -T postgres \
  psql -U postgres -d vecadvisor < examples/demo.sql

The default test DSN is:

postgresql://postgres:postgres@localhost:5432/vecadvisor

vecadvisor analyze \
  --dsn postgresql://postgres:postgres@localhost:5432/vecadvisor \
  --table public.documents \
  --vector embedding

This reports table rows, vector dimension, indexes, extended statistics, catalog fingerprints, stats freshness, and pgvector capabilities.

Use explain when you have one concrete query vector and want a one-query diagnostic:

vecadvisor explain \
  --dsn postgresql://postgres:postgres@localhost:5432/vecadvisor \
  --table public.documents \
  --vector embedding \
  --query "tenant_id = 1" \
  --q-vector examples/query-vector.json \
  --probe-rows 16 \
  --format text

explain reads --q-vector, runs one local selectivity probe, observes the planner's vector query shape, and renders an EXPLAIN VECTOR style report.

Use recommend when you have representative query vectors for a workload:

vecadvisor recommend \
  --dsn postgresql://postgres:postgres@localhost:5432/vecadvisor \
  --table public.documents \
  --vector embedding \
  --query "tenant_id = 1" \
  --q-vectors examples/query-vectors.json \
  --probe-rows 16 \
  --max-query-vectors 3 \
  --local-cache-dir .vecadvisor-cache

recommend uses p10 local selectivity across the vector sample. It emits:

• Ranked candidate strategies.
• Estimated latency, recall, returns-k, and confidence.
• A concise decision summary.
• PostgreSQL observed plan, when a vector sample is available.
• Statistics suggestions such as CREATE STATISTICS.
• Optional partial index or partition strategy suggestions.

If you do not have production query vectors yet, you may explicitly opt in to a low-confidence table-sample fallback:

vecadvisor recommend \
  --dsn postgresql://postgres:postgres@localhost:5432/vecadvisor \
  --table public.documents \
  --vector embedding \
  --query "tenant_id = 1" \
  --allow-table-sample-vectors

Table-sampled vectors are marked in the output and reduce confidence.

Repeated multi-vector probes can be cached:

vecadvisor recommend ... --local-cache-dir .vecadvisor-cache

Cache keys include:

• table name
• filter shape
• probe rows
• query vector source
• query vector fingerprint
• table statistics fingerprint
• index fingerprint

Changing table statistics or index definitions naturally produces a different cache key. Cache entries contain only aggregate local selectivity, not raw vectors or row ids.

Use --refresh-local-cache to ignore and replace an existing cache entry.

Synthetic calibration fits cost constants and an unfiltered recall curve:

vecadvisor calibrate \
  --dataset synthetic \
  --rows 5000 \
  --dim 64 \
  --queries 50 \
  --clusters 16 \
  --filter-selectivity 0.1 \
  --correlation 0.5 \
  --out profiles/local.json

Load a profile during explain or recommend:

vecadvisor recommend ... --calibration profiles/local.json

Without a profile, the advisor uses conservative default constants and marks that in candidate notes.

Run a small synthetic benchmark:

vecadvisor benchmark \
  --dataset synthetic \
  --strategies exact,postfilter,iterative \
  --rows 512 \
  --dim 16 \
  --queries 8 \
  --clusters 8 \
  --filter-selectivity 0.1 \
  --correlation 0.5 \
  --limit 10 \
  --ef-search 40 \
  --out benchmarks/synthetic.json

Run a selectivity and correlation sweep:

vecadvisor benchmark-sweep \
  --dataset synthetic \
  --rows 512 \
  --dim 16 \
  --queries 8 \
  --clusters 8 \
  --filter-selectivities 0.01,0.05,0.1 \
  --correlations 0,0.5 \
  --limit 10 \
  --ef-search 40 \
  --out benchmarks/sweep.json

Analyze crossover behavior:

vecadvisor crossover benchmarks/sweep.json --out benchmarks/crossover.json

Build a publishability proof report:

vecadvisor proof benchmarks/sweep.json --out benchmarks/proof.json

Render a crossover chart:

vecadvisor plot-crossover benchmarks/sweep.json --out benchmarks/crossover.svg

Important JSON fields:

• catalog_snapshot: stats and index fingerprints for reproducibility.
• selectivity: advisor-vs-PostgreSQL selectivity cross-check with severity.
• local_selectivity: p10 and median local selectivity, confidence, and notes.
• local_selectivity_cache: cache hit/store metadata when enabled.
• recommendation.decision: winner, runner-up, viability, confidence, and why.
• recommendation.ranked: full candidate plan and estimate details.

• Alpha quality: APIs and output shape may change before v0.1.
• The current release is an external CLI. It does not change PostgreSQL planner behavior.
• The SIFT1M benchmark commits JSON/SVG results only; source vectors are downloaded into ignored local data/ paths.
• Calibration constants are workload and hardware dependent.
• Predicate parsing intentionally supports a restricted safe subset of filters.
• Native exact top-k ground truth is available when a locally built shared library is configured; prebuilt native wheels are deferred until the ABI and CI matrix are stable.
• Postgres integration tests require a reachable PostgreSQL database with pgvector installed. Tests skip those cases if the database is unavailable.

Install dev dependencies:

python -m pip install -e ".[dev]"

Run checks:

python -m pytest
python -m ruff check .
python -m mypy src/vecadvisor
python -m build

Build the optional native distance kernels when a C++17 compiler and CMake are available:

cmake -S native -B native/build
cmake --build native/build --config Release --parallel
ctest --test-dir native/build --output-on-failure

Run the native distance benchmark wrapper to compare scalar and runtime-dispatch kernels against deterministic NumPy checksums:

python tools/native_distance_benchmark.py \
  --rows 4096 \
  --queries 16 \
  --dim 128 \
  --iterations 5

It writes a JSON artifact, Markdown report, and SVG chart under docs/.

The committed MVP2 native artifact compares AVX2 dispatch with an explicit scalar-only fallback build and the optional SimSIMD external baseline: docs/benchmarks/native-distance-kernels.md. The native C ABI also includes a bounded top-k helper for exact ground-truth search over row-major float32 blocks without materializing an N x Q distance matrix.

Python benchmark ground-truth can use that helper when VECADVISOR_NATIVE_DISTANCE_LIB points at the built shared library. Runs record ground_truth.native_used; if the library is missing, VecAdvisor falls back to the NumPy path.

This project builds from public PostgreSQL, pgvector, and filtered ANN literature. It does not copy source from AGPL or source-available vector extensions. Ideas such as local selectivity, filtered ANN recall risk, iterative scans, partial indexes, and calibration are implemented from public documentation, papers, and first-principles cost modeling.

Apache License 2.0. See LICENSE.

Built and maintained by Varun Sharma.