ipo_drafting_agent

DRHP Drafting Agent – Capital Structure & Shareholding Pattern

AI agent that drafts the Capital Structure and Shareholding Pattern section of a Draft Red Herring Prospectus (DRHP) from ROC filings and supporting documents.

Features

• RAG: Ingest PAS-3, SH-7, List of Allottees, MOA from data/ (markdown)
• Agentic retrieval: Multi-query search via search_knowledge
• Structured output: @structured DraftOutput (Syrin schema, OpenAI-compatible)
• Automation transparency: Sources used, auto-extracted parts, items needing review
• Grounding: Fact extraction and verification from source documents
• Guardrails: FactVerificationGuardrail to verify grounded facts
• Budget: Cost limits per run
• Output: Markdown (.md) and optional DOCX
• HTTP serving: serve.py for /chat and /playground

Requirements

• Python 3.11+ (syrin)
• OPENAI_API_KEY

Setup

From the syrin-python repo root:

cd examples/ipo_drafting_agent
uv sync  # or: uv pip install -e "../..[knowledge,docx,serve]" python-dotenv

Create .env with OPENAI_API_KEY, or export it.

Run

# From ipo_drafting_agent (or repo root with PYTHONPATH)
python run.py

Output: output_draft_section.md (draft + transparency metadata). Use --trace to inspect intermediate steps.

Serve over HTTP

python serve.py

Then:

curl -X POST http://localhost:8000/chat \
  -H "Content-Type: application/json" \
  -d '{"message": "Draft the Capital Structure and Shareholding Pattern section."}'

Or visit http://localhost:8000/playground for the web UI.

Output Format

• output_draft_section.md – draft section + automation transparency (sources, auto-extracted, requires review)
• output_draft_section.docx – same as Word (if syrin[docx] installed)

Intermediate structured data is visible via --trace; no separate JSON file.

Input Documents

Add documents under data/:

• data/PAS-3/ – PAS-3 forms, allotment resolutions, list of allottees
• data/SH-7/ – SH-7, MOA, board resolutions, EGM notices

Syrin Components Used

Component	Purpose
Knowledge	Document loaders, chunking, embedding, vector store
Agent	LLM + tools, structured output
search_knowledge	RAG retrieval (auto-added via knowledge=)
Output(DraftOutput)	@structured output
GroundingConfig	Fact extraction and verification
FactVerificationGuardrail	Verify grounded facts
Budget	Cost limits per run

Tests

PYTHONPATH=examples/ipo_drafting_agent uv run pytest examples/ipo_drafting_agent/ipo_tests/ -v