RAG Dataset Linter

Audit already-chunked RAG datasets before paying to embed them or sending them into a vector database.

RAG Dataset Linter checks Apify Dataset rows, Actor Run datasets, or inline records for structural defects and retrieval-quality risks. It produces one decision-ready summary, a separate findings dataset, a JSON report, and an optional Markdown report.

What it catches

• empty, missing, undersized, and oversized chunk text
• exact and near-duplicate chunks
• duplicate chunk IDs
• missing or duplicate chunk indexes
• gaps in per-document chunk sequences
• excessive overlap between adjacent chunks
• missing document and source provenance
• repeated boilerplate across documents
• leaked HTML markup
• substantial chunks without title or heading context
• contextless references such as “shown above”

Structural problems are errors. Conservative linguistic checks are warnings only. Issue codes and evidence are the product; the gate result is a compact automation signal.

Input

Provide exactly one source:

• datasetId for an Apify Dataset selected with read-only resource access
• runId for an accessible Actor Run's default dataset
• records for a small inline sample

When the Actor is added as an integration, it also accepts Apify's implicit payload.resource.defaultDatasetId. An explicit datasetId takes precedence over that fallback. Selecting a Dataset grants only read access to that resource; the Actor remains limited-permission and never receives account-wide access.

Resolving runId requires access to that Run's metadata. For private Run output, select its default Dataset with datasetId or add this Actor as a Dataset integration so Apify can grant the required read-only resource access.

The Actor auto-detects common fields such as chunkText, text, chunkId, documentId, url, chunkIndex, headingPath, and metadata.title. Use fieldMapping dot-path overrides for other schemas.

{
  "datasetId": "YOUR_DATASET_ID",
  "maxItems": 1000,
  "fieldMapping": {
    "text": "chunkText",
    "chunkId": "chunkId",
    "sourceUrl": "url",
    "chunkIndex": "chunkIndex"
  },
  "minTokens": 40,
  "maxTokens": 2000,
  "nearDuplicateSimilarity": 0.95,
  "maxAdjacentOverlapRatio": 0.35,
  "failRunOnError": false
}

Outputs

The default dataset contains exactly one audit_summary row with:

• PASS, WARN, or FAIL
• recommended automation action: continue, review, or stop
• item and finding counts
• issue-code counts
• exact and near-duplicate rates
• token-size percentiles
• metadata coverage
• estimated embedding tokens wasted by exact duplication and excess overlap
• applied limit and truncation warnings

The findings dataset contains one row per stored finding with a stable ID, source record index, chunk context, code, severity, message, bounded evidence preview, and related-record reference.

Aggregate JSON is stored under OUTPUT. When enabled, a human-readable report is stored under REPORT.md.

Automation gate

• PASS: no findings
• WARN: warning findings only
• FAIL: at least one structural error

Set failRunOnError to true to write every output and then fail the Actor run when the gate is FAIL. This allows schedules and workflow tools to stop bad data before embedding or indexing it.

Limits and safety

• maximum 25,000 audited records
• maximum 50 MB of mapped chunk text
• maximum 20,000 stored findings
• inline input maximum 1,000 records
• fixed 512 MB Actor memory
• evidence previews are bounded; complete source records are never copied into reports
• selected private datasets use an explicit read-only resource grant while the Actor remains limited-permission

The Actor does not crawl websites, call external URLs, create embeddings, use LLMs, mutate the source dataset, write to vector databases, detect contradictions, or produce a repaired corpus.

Pricing

The Actor charges one custom PPE event, chunk-audited, per source record admitted for analysis. Platform usage is included in the event price.

• Free: $1.50 per 1,000 audited chunks
• Bronze: $1.00 per 1,000 audited chunks
• Silver: $0.75 per 1,000 audited chunks
• Gold and above: $0.50 per 1,000 audited chunks