PDF-RAG API Skill

Use this skill when: curl/jq fails, response shape is wrong, or you're unsure of the endpoint.

Cheat Sheet

Response Paths

Stage Names

extract_parties, extract_dates, extract_land_description, extract_conveyance, extract_document_classification
markdown, segmentation, assembly, runsheet, runtime_linking

Shell Command Pattern

IMPORTANT: Never use shell variable interpolation in curl commands. Always inline values directly:

# WRONG - variable interpolation can fail
DOC_ID="abc123" && curl -s "http://localhost:8000/api/documents/$DOC_ID/parties"

# RIGHT - inline the value directly
curl -s "http://localhost:8000/api/documents/abc123/parties"

Gotchas & Common Mistakes

Estate Graph (L3)

• GET vs POST: GET retrieves stored graph, POST rebuilds from L0-L2 data
• estate_type is lowercase: "fee", "mineral", "leasehold" (not uppercase)
• Chain = nodes grouped by tract: A "mineral chain" = all mineral nodes sharing same tract_id
• Quick summary: .summary.mineral_estates gives count
• List nodes by type: .dag.nodes[] | select(.estate_type == "mineral")
• Group into chains: [.dag.nodes[] | select(.estate_type == "mineral")] | group_by(.tract_id)
• Count chains: Above query | length

Rebuild Estate Graph (Exploratory Mode 1.1)

Endpoint: POST /api/packages/{package_id}/estate-graph

Request body (JSON):

{"version": "1.1"}

Full curl command:

curl -s -X POST "http://localhost:8000/api/packages/{package_id}/estate-graph" \
  -H "Content-Type: application/json" \
  -d '{"version": "1.1"}'

Version options:

• "1.0": Exploratory Layers - POC 5-layer fuzzy matcher (L0-L4 gates)
• "1.1": Exploratory Signals - EdgeScore-based matcher (includes comparison field)
• "2.0": Operative deterministic matcher (exact matches only)

Response paths:

• Nodes: .dag.nodes[]
• Edges: .dag.edges[]
• Gaps: .dag.gaps[]
• Summary: .summary
• Exploratory outcomes (v1.1): .exploratory_outcomes
• Winner/runner-up comparison (v1.1): .exploratory_outcomes[estate_id].comparison

Comparison data (v1.1):

.exploratory_outcomes[estate_id].comparison:
  .winner_id: string
  .runner_up_id: string | null
  .rank_delta: number
  .top_winner_reasons[]: {signal, winner_state, runner_up_state, differentiation}
  .top_runner_up_losses[]: {signal, issue}
  .suppression_analysis: {eligible, blocking_contradictions[]}

Chain Analysis (deprecated - use Estate Graph)

• Use the endpoint, not Package.meta: Chain analysis results are in GET /packages/{id}/chain-analysis, not buried in Package.meta.chain_analysis
• L2 before L3: Run L2 clustering (POST /packages/{id}/l2/parties) BEFORE chain analysis for proper cross-document party matching
• Chain types are dict keys: Access via .chains["fee"], .chains["mineral"], .chains["leasehold"]

Extraction Endpoints

• Always use .resolved: The clean data is in .resolved, raw LLM output is in .extraction (debug mode only)
• Debug mode: Add ?debug=true to get raw extraction + normalized + extraction_context
• Party IDs are document-scoped: party_id is unique within a document, use L2 clusters for cross-doc identity

List Endpoints

• Documents list: .items[] (paginated response)
• Packages list: .packages[] (NOT .[])
• Package documents: .documents[] (nested in package detail)
• Chain edges: .chains["fee"].dag.edges[]

Processing

• Stages array: POST /api/process takes stages: ["extract_parties", "extract_dates", ...]
• Poll for completion: GET /api/jobs/{job_id} until status is "completed" or "failed"
• enrich_only=true: Re-run L1 enrichment without re-running L0 extraction (saves LLM cost)

Need full schema details? Read .claude/skills/api-reference/API_REFERENCE.md