docs

Imported from yashverma9/AI-Augmented-Investment-Pipeline-Emergence (docs/AGENTS.md). Install upstream with npx skills add yashverma9/AI-Augmented-Investment-Pipeline-Emergence --skill docs. Copyright stays with the author.

Project Instructions

Stack

• Python, managed with uv (pyproject.toml + uv.lock)
• langchain_openai.ChatOpenAI as a thin structured-output LLM client only — no LangGraph, chains, or agents
• pydantic for schemas, tenacity for retries, httpx for async HTTP, jinja2 for memo templates

Architecture

• pipeline is the root directory for the python code
• CLI entrypoint: run.py (--topic, --stage source|analyze|memo|all)
• Three decoupled, file-based stages — each reads a file and writes a file, never calls another stage's functions directly:
  • sourcing.py -> candidates.json
  • analysis.py -> analysis.json (reads candidates.json)
  • memo.py -> memos/*.md (reads analysis.json)
• Shared modules: models.py (pydantic schemas), thesis.py (criteria/weights/thresholds), llm.py (LLM client + retry wrapper), product_hunt.py (sourcing API client)
• Raw API responses are cached to disk under cache/ (gitignored) before parsing
• Outputs (candidates.json, analysis.json, memos/*.md) are committed to the repo — not gitignored

Rules

• Prefer existing modules and utilities.
• Do not introduce new dependencies unless necessary.
• Do not modify unrelated files.
• Do not rewrite working code without a reason.
• Keep implementations simple and production-ready.
• Do not change the pydantic schemas (models.py) unless explicitly asked.
• Structured LLM output must be validated against a pydantic model — never parse freeform text with regex.
• Never guess-fill missing data; mark it null with a basis/explanation instead.
• Deterministic logic (scores, thresholds, Pass/Watch/Meeting calls) stays in code, not another LLM call.

Before finishing a task

• Run the relevant tests.
• Run lint/typecheck when applicable.
• Fix errors caused by your changes.
• Summarize what changed and what was tested.