SEC 10-K Verifiable Citation & Financial Analysis Copilot Playbook
End-to-end 6-stage AI scenario playbook for SEC 10-K equity analysis: analyst ROI framing, Docling table-faithful ingestion, LangGraph deterministic math routing, inline citation HITL UX, and RAGAS numerical faithfulness CI gates.
Best For
Research workflows reconciling multi-year gross margins, cash flows, and Risk Factors across 10-K filings with clickable page-coordinate evidence verification.
Human-AI Boundary
Split-pane UI pairs side-by-side PDF evidence bounding boxes with clickable inline `[Filing-Page-Table]` anchors; analyst approval is mandatory whenever unit scales or fiscal restatements conflict.
Anti-Patterns to Avoid
- Building a naive Chat-with-PDF demo using fixed 500-token chunks and asking the LLM to compute YoY growth percentages mentally
- Shipping without period-mismatch regression tests (e.g., answering Q3 YTD numbers when asked for full-year FY metrics)
Latency/Cost Profile & Eval Metrics
Streaming TTFT < 1.2s, full verified table comparison 2.5-4.5s; Offline per-filing ingestion < $0.08 · Reranker compute + prompt context size (offset by semantic cache & table JSON compression)
- Table Cell Provenance F1 & Numerical Exact Match: >= 0.96 on 100-question SEC 10-K benchmark (Zero tolerance for LLM mental math discrepancy against Python calculator output)
- RAG Triad Context Recall & Faithfulness: Recall >= 0.92, Faithfulness >= 0.96 (Executed on every prompt/chunking PR via automated offline eval gate)
Bidirectional Taxonomy Links
Internal Tools & Primary Sources
- SEC EDGAR Full-Text Filing Corpus & Taxonomy Standards (sec_filing, 2026-11-15)
- Docling Open-Source Financial Table Parsing (github, 2026-11-15)
- LangGraph Stateful Agent Orchestration Documentation (official_docs, 2026-11-15)