Pattern: Batch intake — file drop → text extraction → AI entity extraction → KB cross-reference → flagged report

Skills used: ai-document-analysis, document-research-specialist, public-records-research-specialist

Workflow overview

[File Trigger / Webhook] → [Text Extraction] → [AI Entity Extraction]
    → [Entity Lookup] → [Confidence Filter] → [Output Flagged Report]

Node chain

1. File Trigger

• Type: Webhook (POST, multipart/form-data) or Local File Trigger (watched folder)
• Path: /webhook/document-triage
• Accepts: PDF, DOCX, images (PNG/JPG for scanned docs)

2. Text Extraction

• Type: HTTP Request to Apache Tika (if running) or Code node with pdf-parse
• URL: http://tika:9998/tika (if Tika container is available)
• Fallback: For images, base64-encode and send to Claude Vision API in step 3 instead.
• Output: Extracted text content per page.

3. AI Entity Extraction

• Type: HTTP Request
• Method: POST
• URL: {{ $vars.PATRIOT_API_URL }}/chat/message
• Body:

  {
    "message": "Extract all named entities (people, organizations, dates, monetary amounts) from the following document text. Return structured JSON with confidence scores per entity.\n\n{{ $json.extracted_text }}",
    "persona": "investigator"
  }

• Purpose: Use Claude to identify entities, relationships, and key facts.
• Output: JSON with entities array and relationships array per the ai-document-analysis schema.

4. Entity Lookup (KB cross-reference)

• Type: HTTP Request (looped over extracted entities)
• Method: POST
• URL: {{ $vars.PATRIOT_API_URL }}/investigate/entity-lookup
• Body per entity:

  {
    "query": "{{ $json.entity_name }}",
    "categories": ["accountability"],
    "limit": 5
  }

• Purpose: Check each extracted entity against the 336+ accountability profiles.

5. Confidence Filter

• Type: IF node
• Condition: Only pass entities with:
• Extraction confidence >= medium, AND
• At least one accountability profile match with relevance_score > 0.3
• Purpose: Reduce noise. Low-confidence extractions go to a review queue.

6. Output Flagged Report

• Type: Respond to Webhook (or write to n8n database / send notification)
• Body: Structured report with:
• Document metadata (filename, page count, extraction method)
• Flagged entities with profile matches
• Low-confidence entities for human review
• Provenance tags on all extracted data (AI-extracted, AI-inferred)

Provenance tagging

Per ITI inferred-data transparency rules, every entity in the output carries:

When to use

• Processing FOIA document productions (10-1000+ pages)
• Triaging leaked or whistleblower documents
• Scanning financial disclosures for accountability profile matches
• Batch processing court filings for entity extraction