HWsqlite

SQLite-backed hybrid RAG index for Huawei Ascend C and CANN documentation.

This repo contains:

• hw_rag.sqlite: prebuilt SQLite index, stored with Git LFS.
• hw_rag.dataset.json: dataset manifest defining the current multi-source corpus.
• hw_rag/: ingestion, classification, chunking, FTS5, embedding, search, and MCP server code.
• hw_rag_cli.py: command-line build/search entrypoint.
• scripts/test_mcp_server.py: MCP smoke test.
• tests/: basic regression tests.
• HW_RAG_README.md: usage details.

The default dataset currently combines:

• docs/asc-devkit
• docs/cann-docs

Quick Search

python hw_rag_cli.py search "DataCopy GM UB LocalTensor" --db hw_rag.sqlite --limit 8
python hw_rag_cli.py search "TQue AllocTensor FreeTensor TPipe" --db hw_rag.sqlite --limit 8
python hw_rag_cli.py search "aclnnAddRmsNorm" --db hw_rag.sqlite --limit 8

MCP

python scripts/test_mcp_server.py

Codex:

codex mcp add ascend-c-rag --env PYTHONIOENCODING=utf-8 --env PYTHONPATH=/absolute/path/to/repo -- python -m hw_rag.mcp_server --db /absolute/path/to/repo/hw_rag.sqlite --embedding-provider hash

Claude Code:

claude mcp add ascend-c-rag -e PYTHONIOENCODING=utf-8 -e PYTHONPATH=/absolute/path/to/repo -- python -m hw_rag.mcp_server --db /absolute/path/to/repo/hw_rag.sqlite --embedding-provider hash

Use an absolute --db path for MCP servers. Some clients do not guarantee the subprocess working directory, and a relative SQLite path can silently point at a different file.

See HW_RAG_README.md for rebuild and integration details.