addon-rag-ingestion-pipeline（addon RAG ingestion pipeline 搜索）

Add-on: Multi-Format RAG Ingestion Pipeline

Use this skill when an existing project needs RAG ingestion/retrieval across multiple document formats.

Compatibility

• Works with architect-python-uv-batch.
• Works with architect-python-uv-fastapi-sqlalchemy.
• Can back a Next.js app via a Python worker service.

Inputs

Collect:

• SOURCE_FORMATS: one or more of pdf, markdown, txt, html, csv.
• EMBED_PROVIDER: openai or sentence-transformers.
• VECTOR_STORE: pgvector, chroma, or existing vector layer.
• CHUNK_SIZE: default 1000.
• CHUNK_OVERLAP: default 150.
• TOP_K: default 5.

Integration Workflow

1. Add dependencies (Python worker path):

uv add pypdf markdown-it-py beautifulsoup4 pandas langchain-text-splitters

• If EMBED_PROVIDER=openai: uv add openai
• If EMBED_PROVIDER=sentence-transformers: uv add sentence-transformers
• If VECTOR_STORE=chroma: uv add chromadb

1. Add modules:

src/{{MODULE_NAME}}/rag/
  loaders/pdf_loader.py
  loaders/markdown_loader.py
  loaders/text_loader.py
  loaders/html_loader.py
  loaders/csv_loader.py
  normalize.py
  chunking.py
  embeddings.py
  indexer.py
  retriever.py

1. Use a normalized document contract:

• document_id
• source_path
• source_type
• content
• metadata (filename/page/section/checksum/ingested_at/model_version)

1. Implement ingestion entrypoint:

uv run {{PROJECT_NAME}} rag-ingest --source ./data/inbox --formats pdf,markdown,txt

1. Implement retrieval entrypoint:

uv run {{PROJECT_NAME}} rag-query --q "question" --top-k 5

• Ensure both commands are wired in the project CLI/script entrypoint.
• rag-query depends on an existing index from rag-ingest; do not run these validation commands in parallel.

Loader Notes

• PDF: extract per page and keep page_number in metadata.
• Markdown: keep heading hierarchy and section anchors in metadata.
• Text: detect encoding fallback (utf-8, then latin-1).
• HTML: strip script/style tags and preserve title/headings where possible.
• CSV: convert rows into stable textual records with row identifiers.

Minimal Defaults

normalize.py

import re
import unicodedata

def normalize_text(raw: str) -> str:
    text = unicodedata.normalize("NFKC", raw)
    text = text.replace("\r\n", "\n")
    text = re.sub(r"[ \t]+", " ", text)
    text = re.sub(r"\n{3,}", "\n\n", text)
    return text.strip()

chunking.py

from langchain_text_splitters import RecursiveCharacterTextSplitter

def chunk_text(text: str, chunk_size: int = 1000, overlap: int = 150) -> list[str]:
    splitter = RecursiveCharacterTextSplitter(
        chunk_size=chunk_size,
        chunk_overlap=overlap,
        separators=["\n\n", "\n", ". ", " ", ""],
    )
    return splitter.split_text(text)

Guardrails

• Documentation contract for generated code:

- Python: write module docstrings and docstrings for public classes, methods, and functions. - Next.js/TypeScript: write JSDoc for exported components, hooks, utilities, and route handlers. - Add concise rationale comments only for non-obvious logic, invariants, or safety constraints. - Apply this contract even when using template snippets below; expand templates as needed.

• Deduplicate ingestion by checksum to keep re-runs idempotent.
• Store embedding model/version so re-indexing can be reasoned about.
• Never interpolate user queries into raw SQL vector search.
• Keep ingestion async/offline for large corpora; do not block request-response paths.
• Preserve citation metadata for retrieval (source_path, section, page, row id).

Validation Checklist

• Confirm generated code includes required docstrings/JSDoc and rationale comments for non-obvious logic.

uv run {{PROJECT_NAME}} rag-ingest --source ./data/inbox --formats pdf,markdown
uv run {{PROJECT_NAME}} rag-query --q "smoke test" --top-k 5
uv run pytest -q

Decision Justification Rule

• Every non-trivial decision must include a concrete justification.
• Capture the alternatives considered and why they were rejected.
• State tradeoffs and residual risks for the chosen option.
• If justification is missing, treat the task as incomplete and surface it as a blocker.