Filing Intelligence RAG — Data skill for Claude Code

filing-intelligence-rag

A local pipeline that reads public financial filings and produces a cited, first-pass credit rationale note.

python -m risk_analytics ingest
python -m risk_analytics ask "As at 30 June 2026, what are APSEZ's Net Gearing and DSCR, and how much covenant headroom remains?"

Output is a Markdown note plus a machine-readable trace of every decision that produced it.

**This is a demonstration artifact, not a rating tool.** It is not a credit rating, not investment advice, and not reviewed by a rating committee. Every note it writes says so.

---

What it actually does

PDF ──► classify page ──► extract by class ──► chunk ──► embed ──► Chroma
          text/table/chart    prose | table       tables never     (local)
          from layout stats   structure           split mid-row
                                                       │
question ──► plan sub-queries ──► search ──► reflect ──┘
                    ▲                          │
                    └──── revise and retry ────┘
                                               │
                              router ──► ≤2 of 4 specialists ──► synthesis ──► note

Five things are worth knowing because they are where the design has opinions:

**Citations are built from stored metadata, never written by the model.** A specialist names chunk IDs; the code resolves those IDs to document and page. A claim citing an ID that was not retrieved is **dropped and counted**, not printed with a caveat. This is the mechanism the whole thing rests on — a fabricated page number in front of a credit professional is the failure that ends the conversation.

**Page classification is free and offline.** Layout statistics from PyMuPDF, no model call. Measured at **20/20 on a 20-page holdout sample labelled after the thresholds were frozen** (`tests/labelled_pages.json`), alongside 19/20 on the tuning split.

**Tables never split mid-row.** A row cut in half puts a figure beside the w