officegrep

officegrep

Full-text search inside Word, PowerPoint, Excel and PDF — for agents.

grep cannot see a single character inside a .docx. officegrep extracts the text, indexes it into one SQLite file, and hands it to any agent over MCP.

$ officegrep search "bending moment"

Found 3 hits:

F:\reports\pole-design.docx  [table p 89]
  design bending moment (load point 0.25 m from tip); deflection limit; test moment = 2× …

F:\notes\composites.md  [line 32]
  … utilization 81%. **test bending moment = 2× the standard test moment** …

中文说明：README.zh-CN.md

Why

• grep / ripgrep only read plain text. .docx is a zip of XML, .xlsx keeps strings in a shared table, .pdf is a compressed stream.
• Generic RAG needs an embedding service, a vector store, and a network round-trip.
• This needs one SQLite file: offline, incremental, milliseconds.

Install

pip install git+https://github.com/Asgoresmilelong/officegrep
# optional PDF support
pip install "officegrep[pdf] @ git+https://github.com/Asgoresmilelong/officegrep"

Runtime deps: lxml, openpyxl. PDF needs optional pypdf.

CLI

officegrep index ~/Documents ~/Downloads     # build / update (incremental)
officegrep search "island reef" --limit 5
officegrep search "wind load" --under ~/Documents
officegrep stats

The index lives in ~/.officegrep/index.db; override with --db or OFFICEGREP_DB.

MCP server

officegrep mcp

Client config (Claude Desktop / Cursor / Codex / any MCP host):

{
  "mcpServers": {
    "officegrep": {
      "command": "officegrep",
      "args": ["mcp", "--db", "/Users/me/.officegrep/index.db"]
    }
  }
}

Tool	What it does
search_docs	Full-text search → file path, locator (p 42, slide 3, Sheet1!12, page 7), highlighted snippet
list_docs	Indexed files with block counts
index_folder	Build / refresh the index (incremental)

The protocol is a hand-written JSON-RPC over stdio — no MCP SDK dependency.

How it works

docx / pptx  →  zipfile + lxml, read the XML directly
xlsx         →  openpyxl (read_only)
pdf          →  pypdf (optional)
txt / md     →  line chunks, line numbers kept

• Every block carries a human-readable locator, because the first question an agent asks is "where?".
• The index uses SQLite FTS5 with the trigram tokenizer: substring matching for both CJK and Latin text, no word segmenter required, ranked by bm25().
• Queries of 1–2 characters cannot match trigrams, so they fall back to a scan ranked by hit count (CJK queries are frequently two characters).
• Incremental: compare (size, mtime) first, hash with SHA-1 only when they differ.

Benchmark on a real engineering corpus:

Metric	Value
Files	43 (docx / pptx / xlsx / md, incl. an 8.6 MB deck)
Text blocks	2,260
First index	2.16 s
Database	1.9 MiB

Limits

• No OCR: scanned PDFs yield nothing.
• Literal search, not semantic — "tower" will not find "pole" unless both appear.
• 1–2 character queries scan the table; slow past a few hundred thousand blocks.
• Indexes body text only: no comments, tracked changes, headers or footers.

License

MIT