verifold

AI DATA PIPELINE INTEGRITY

Trust the data
behind your AI.

Know whether your data actually made it into AI search.

Verifold compares evidence across your data pipeline to help you investigate missing updates, retained deletions and retrieval gaps.

Request an engineering pilot Explore the evidence

Developer alpha. Read-only verification. Explicit uncertainty.

EVIDENCE / M2Recorded local result
A document’s journey

Indexed. But not returned.

A serving-route fault left the index intact while the declared query missed its expected result.

  1. Source & captureExpected version observed
    MATCH
  2. Extraction & processingReference artifacts matched
    MATCH
  3. IndexExpected records present
    MATCH
  4. RetrievalExpected top-1 result absent
    GAP
GLOBAL VERDICTUNKNOWN

Scoped observations, not a shared system snapshot.

Built around the pipeline you already have.

PostgreSQLQdrantLlamaIndex local experiment

01 / THE APPROACH

A successful job isn’t
the whole story.

An ingestion run can finish while an old version remains indexed or a serving route misses a document. Verifold makes the observed difference inspectable.

01

Declare what should exist.

Identify the source versions, retained deletions, expected chunks and the scope you intend to verify.

02

Read what actually exists.

Compare source and index metadata. In the supported local pipeline, inspect retained stage artifacts and run explicit retrieval probes.

03

Keep the evidence.

Inspect affected documents, versions, observed gaps and recurrence in local reports—with coverage and uncertainty kept separate.

02 / DEMONSTRATED, WITH LIMITS

Follow the evidence.

Two bounded experiments show what independent verification can reveal. Neither establishes general customer-pipeline support.

M2 / CONTROLLED PIPELINE

The index was complete.
The query still missed.

100 local HTML pages. 300 expected points. A controlled serving exclusion preserved indexed records but broke one declared retrieval expectation.

Read the M2 result
M3 / EXTERNAL STACK

The retry succeeded.
The deleted page remained.

In a four-URL LlamaIndex experiment, a rejected deletion and persisted restart left an old point searchable after retry. Verifold observed the tombstone/index mismatch.

Read the M3 result

03 / PRODUCT SCOPE

A broad direction.
An explicit starting point.

Our vision is end-to-end integrity verification from original sources through crawling, ingestion, processing, indexing and retrieval.

Implemented · developer alpha

Inspect observed integrity.

  • PostgreSQL → Qdrant metadata comparison
  • Online mismatch detection with versions, age and recurrence
  • Local JSON reports and a private local report explorer
  • One controlled crawl-to-search reference pipeline
  • One bounded LlamaIndex deletion/recovery experiment

During live ingestion, the global verdict remains UNKNOWN. Repeated observations do not prove synchronization.

Roadmap · not general shipping support

Verify more of the journey.

  • Source freshness and crawl completeness
  • Transformation and chunk lineage across customer pipelines
  • Embedding and version provenance
  • Retrieval-serving verification across routes and connectors

These directions require new contracts, adapters and independent evidence. No availability dates or production guarantees are implied.

SOURCE-TO-RETRIEVAL VERIFICATION

Start with what
you can prove.

See the observations, the failure conditions and the boundaries of each result.

Request an engineering pilot