filechat-rag（filechat RAG 搜索）

FileChat RAG Skill

Your personal RAG (Retrieval-Augmented Generation) document library backed by Google Drive. Supports multiple Google Drive folders dynamically, interactive folder routing, incremental sync, choosing between Gemini or OpenAI for embeddings, and connecting to Qdrant.

Setup & Bootstrap

FIRST verify that the required environment variables are set in /workspace/skills/filechat/.env:

1. EMBEDDING_PROVIDER (either gemini or openai)
2. GEMINI_API_KEY or OPENAI_API_KEY
3. Optional: QDRANT_URL and QDRANT_API_KEY (If absent, it uses local disk-based JSON).

Create the .env file like this:

echo "EMBEDDING_PROVIDER=gemini" > ./skills/filechat/.env
echo "GEMINI_API_KEY=your_key_here" >> ./skills/filechat/.env

Google Workspace Authentication: Before running any commands, check if the system is authenticated by running:

npx @googleworkspace/cli auth status

If it returns an auth error or indicates no token, you MUST prompt the user to authenticate. Trigger the interactive login flow:

npx @googleworkspace/cli auth login --services drive

Wait for the user to complete the browser OAuth flow before proceeding.

Folder Management

The user can have infinite folders synced. You manage them using folders.js.

• List Folders: cd ./skills/filechat && node folders.js list
• Add a Folder: node folders.js add "Taxes 2026" <FOLDER_ID> (Auto-discovers the ID via gws drive files list if you don't know it!)
• Set Default Folder: node folders.js default "Taxes 2026"

If the user asks to do something with a file/folder but doesn't specify which one, run node folders.js get-default to find the default ID. If no folders exist, ask them to set one up!

How to Sync the Library

When the user asks to "sync", "flush", or "update", you must run the ingestion script.

To sync a specific folder:

cd ./skills/filechat && node sync.js <FOLDER_ID>

To sync EVERYTHING (all folders in the registry):

cd ./skills/filechat && node sync-all.js

*Note: Syncs are highly incremental and use a local cache! If a file hasn't been modified in Drive, the script will skip it instantly and output "0 chunks" embedded. This is NORMAL behavior. If you are debugging, testing, or the user specifically requests a hard flush, you MUST delete the cache files first:*

rm ./skills/filechat/meta_<FOLDER_ID>.json
rm ./skills/filechat/vector_db_<FOLDER_ID>.json

How to Answer User Questions (RAG)

Query the local vector store or Qdrant for the target Folder ID to fetch relevant text chunks:

cd ./skills/filechat && node query.js <FOLDER_ID> "What does my medical discharge say?"

Use the snippets returned to answer the user.

How to Retrieve and Send a Physical File

Find the File ID using the query script, then download it:

gws drive files get --params '{"fileId": "<FILE_ID>", "alt": "media"}' --output /workspace/discharge.pdf

Reply using the media tag: MEDIA:/workspace/discharge.pdf.

How to Store a New File for the User

If the user uploads a file and asks you to save it (or implicitly sends a file per your automatic processing rules):

1. Check their folders (node folders.js list).
2. If they didn't specify which folder, use the default folder. If no default is set, ask them!
3. Notify the user exactly which folder the file is being saved to.
4. Tell the user that you are now extracting the information and saving it in a vectordb.
5. If the file is an image or scanned document, make sure to extract the text using a vision model or OCR before it is embedded. (The sync script handles this natively).
6. Upload it to the correct folder using gws:

   gws drive files create --json '{"name": "filename.pdf", "parents": ["<FOLDER_ID>"]}' --upload /path/to/uploaded/file.pdf

1. Trigger node sync.js <FOLDER_ID> so the vector database chunks and embeds the file into the corresponding vectordb.

How to Test & Validate the Skill

If the user asks you to verify the skill is working, or if you just set it up and want to ensure end-to-end functionality, follow these exact steps:

1. Verify Auth: Run npx @googleworkspace/cli auth status. Ensure it shows a valid token.
2. Verify Drive Access: Do a dry-run fetch of the target folder to ensure GWS can see the files.

   npx @googleworkspace/cli drive files list --params '{"q": "'\''<FOLDER_ID>'\'' in parents and trashed = false"}'

*(If this fails, check folder permissions or GWS credentials.)*

1. Force a Clean Sync: Clear the cache for the test folder to guarantee a fresh run, then sync.

   rm -f ./skills/filechat/meta_<FOLDER_ID>.json ./skills/filechat/vector_db_<FOLDER_ID>.json
   node ./skills/filechat/sync.js <FOLDER_ID>

*(You should see files being downloaded, OCR'd, and chunks being embedded. If it says "0 chunks", verify the folder isn't empty.)*

1. Test the Vector Query: Run a generic query to verify the embeddings were saved and cosine similarity works.

   node ./skills/filechat/query.js <FOLDER_ID> "hello"

*(You should see a list of "Top matches" with similarity scores and text snippets. If you do, the RAG pipeline is 100% operational!)*