| name | wiki-raw-drive |
| description | Fetch and download RAW sources (statements, invoices, documents) from a single Google Drive folder into the local raw cache, without versioning PII in git. Centralizes raw in Drive and keeps the repo lightweight. |
Wiki Raw Drive
Why
Raw sources (bank statements, invoices, PDFs with PII) must NOT be versioned
in git: they leak sensitive data into history and bloat the repo. They live in a
single Drive folder and are downloaded on demand into the local raw cache
(data/raw/, which is gitignored). This separation follows the
kit's architecture: the deterministic code processes; the agent that runs the repo
uses the Drive connector to gather the sources.
Registry
The filename -> drive_file_id map lives in
data/raw/_drive_manifest.json (this JSON is
versioned; it contains no PII, only identifiers). The canonical Drive folder is
configured by the repo owner and referenced by folder id in the manifest.
Flow (executed by the agent that runs the repo)
- To ingest a raw source, find the
drive_file_id in the manifest (or
search the folder via the Drive connector: search_files with parentId = '<folder_id>').
- Download the content with the Drive connector (
download_file_content /
read_file_content) into the raw cache (data/raw/,
gitignored).
- Run the deterministic pipeline over the local file (manifest, text, chunks,
index) and follow with the contextual LLM pass (skill
wiki-llm-context-agent).
- If you download a new file, register
filename -> drive_file_id in the manifest.
Rules
- Never commit raw files; only the
_drive_manifest.json is versioned.
- Never paste PII content (account number, CPF, transactions) into versioned
pages; private pages may summarize, public pages require redaction and a
promotion gate.
- If permission is missing on the Drive connector, ask the user to authenticate.
- The auditor blocks secrets in any file and PII in public pages;
run it before consolidating.