vhecfsck

Read-only, empirical, offline auditor for vector indexes that detects silent recall decay and index pathologies before they reach production.

The name is a personalized play on Unix fsck: a vector-index checker, with an H from the author's name, Hector. Identifiers stay lowercase vhecfsck.

Quickstart

Run the interactive CLI demonstration from any machine with Python ≥ 3.11:

uvx vhecfsck demo

Or from a local checkout:

uv run vhecfsck demo

---

The Problem

Vector database dashboards show green HTTP status and low latency while search quality silently degrades due to unindexed deletions, centroid drift, or graph fragmentation.

Production workloads encounter silent recall decay across major vector engines, as documented in Qdrant #7147 (tombstone accumulation degrading search precision), pgvector #244 (HNSW graph disconnected components post-deletion), Lance #4164 (IVF partition centroid drift after out-of-order appends), and table-level tombstone accumulation under default IVFFlat lists.

vhecfsck executes empirical audit queries against exact ground truth to detect and quantify these silent pathologies without modifying target database state.

Explicit Limitations

• 3D Projection: Spatial visualizer output is a lossy projection for visual intuition, not an exact spatial distance metric.
• Corpus-Drawn Queries: Recall bounds derived from corpus vectors represent an optimistic upper bound compared to external query distributions.
• Hubness Thresholds: Hub share and antihub fraction metrics depend on dimensionality $d$ and sample size $|S|$ (calibrated via threshold profiles).
• pgvector DFI: Deletion Fragmentation Index on PostgreSQL is a table-level proxy based on n_dead_tup statistics rather than exact per-segment tombstone inspection.

---

What the Tool Measures

vhecfsck measures five core index health metrics against exact ground truth calculations:

Metric	Target Pathology	Warn Threshold	Fail Threshold	Direction
Canary Recall	Silent recall decay	< 0.85	< 0.70	Lower is worse
Hub Share (top 1%)	Hubness / central point dominance	> 0.20	> 0.35	Higher is worse
Antihub Fraction	Orphaning / unreachable vectors	> 0.25	> 0.40	Higher is worse
Deletion Fragmentation Index (DFI)	Tombstone accumulation	> 0.15	> 0.30	Higher is worse
Partition Size CV	IVF centroid imbalance / skew	> 1.20	> 2.00	Higher is worse

For full metric mathematical formulations, see roadmap/02-metrics-spec.md.

---

CI Integration & Exit Codes

Integrate vhecfsck into CI/CD pipelines to block deployments when index quality degrades.

Exit Codes

Exit Code	Label	Description
0	OK	All enabled metrics pass thresholds.
1	WARN	One or more metrics breached warning thresholds.
2	FAIL	One or more metrics breached failure thresholds.
3	INCONCLUSIVE	Audit cannot determine state (unsupported capability or insufficient sample size).
4	USAGE	CLI argument error, target connection failure (TargetConnectionError), or memory budget exceeded (ResourceError).
70	INTERNAL	Unhandled internal error.

CI Recipe (GitHub Actions)

name: Audit Vector Index

on:
  schedule:
    - cron: '0 2 * * *'
  workflow_dispatch:

jobs:
  audit:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/setup-python@v5
        with:
          python-version: '3.11'
      - name: Install uv
        run: curl -LsSf https://astral.sh/uv/install.sh | sh
      - name: Run Index Audit
        run: |
          uvx vhecfsck audit postgres://user:pass@localhost:5432/vectors \
            --only canary_recall,dfi \
            --format text

For GitLab CI, Kubernetes CronJob, and Prometheus textfile recipes, see CI Integration Guide.

---

Engine Capability Matrix

Different target engines expose different level of introspection capabilities. Unsupported capabilities return UNAVAILABLE (exit code 3) rather than fake values.

Capability	Synthetic	LanceDB / Lance	Qdrant	Postgres / pgvector
Exact k-NN Ground Truth	Yes	Yes (native / compute)	Yes (brute force)	Yes (sequential scan)
Canary Recall	Yes	Yes	Yes	Yes
Hubness Analysis	Yes	Yes	Yes	Yes
DFI (Tombstones)	Yes	Yes (exact _rowid)	Partial (segment telemetry)	Proxy (n_dead_tup)
Partition Size CV	Yes	Yes (IVF lists)	UNAVAILABLE	Partial (catalog stats)

See Consolidated Capability Matrix and individual engine guides (LanceDB, Qdrant, pgvector).

---

Read-Only Guarantee & Zero Egress

vhecfsck is 100% read-only and designed for secure, air-gapped environments:

• Zero Writes: Never executes VACUUM, REINDEX, INSERT, UPDATE, or DELETE.
• Zero Network Egress: No telemetry, analytics, tracking, or external asset downloads.
• Database Isolation: PostgreSQL sessions enforce default_transaction_read_only=on and Connection.read_only = True.
• Snapshot Assurance: File-backed engines (LanceDB) verify zero disk modifications via SHA-256 snapshots and chmod -R a-w read-only mounts.

See Read-Only Assurance and SECURITY.md.

---

Measured Performance

All published numbers are measured on designated reference hardware (Apple Silicon 8-core CPU, macOS 26.5, Python 3.11.15, Apple Accelerate BLAS):

Stage / Component	Input Scale	Measured Duration	Peak RSS	Budget Ceiling	Status
Ground Truth (exact_knn)	$100,000 \times 768$	0.6974 s	1,832.22 MB	5.0 s	Pass
Ground Truth (exact_knn)	$1,000,000 \times 768$	5.7620 s	1,861.86 MB	20.0 s	Pass
Hubness Subsample ($S=20k$)	$20,000 \times 768$	0.6125 s	1,980.20 MB	3.0 s	Pass
Deterministic 3D Projection	$1,000,000 \times 768$	0.1611 s	4,323.00 MB	2.0 s	Pass
Full Audit End-to-End	$100,000 \times 768$	0.2045 s	4,323.00 MB	5.0 s	Pass

See docs/performance.md for benchmark instructions.

---

Installation & Extras

Quickstart

uvx vhecfsck demo

Package Installation

pip install vhecfsck

Engine Extras

pip install "vhecfsck[lancedb]"
pip install "vhecfsck[qdrant]"
pip install "vhecfsck[postgres]"
pip install "vhecfsck[server]"

Or using uv:

uv add "vhecfsck[lancedb,qdrant,postgres]"

---

Documentation & References

• Metrics Specification: roadmap/02-metrics-spec.md
• Read-Only Assurance: docs/read-only.md
• Performance Guidance: docs/performance.md
• Calibration Data: docs/calibration/README.md
• Contributing Panel: CONTRIBUTING.md

---

Licence & Credit

Copyright © hbauzan. Published under the Apache License 2.0. See NOTICE for details.