Glean Chatbot

Indexes a small set of internal contracts into Glean, answers questions from those docs via Search + Chat, and exposes the same flow as an MCP tool (ask_glean).

Requirements

• Python 3.10+
• A Glean instance
• An Indexing API token
• A Client API token with Search and Chat scopes

Setup

1. Clone and enter the repo:

git clone <your-repo-url>
cd glean-chatbot

2. Install dependencies:

pip install -r requirements.txt

3. Create a .env from the example and fill in your values:

# macOS / Linux
cp .env.example .env

# Windows (PowerShell)
copy .env.example .env

Variable	Required	Purpose
GLEAN_INSTANCE	yes	Instance name (https://<instance>-be.glean.com)
GLEAN_INDEXING_TOKEN	yes	Indexing API token
GLEAN_CLIENT_TOKEN	yes	Client API token (Search + Chat)
GLEAN_ACT_AS	yes*	User email for X-Glean-ActAs (*needed with a global client token)
GLEAN_DATASOURCE	no	Datasource name (default interviewds)
GLEAN_DATASOURCE_DISPLAY	no	Display name in the Glean UI
GLEAN_URL_REGEX	no	Regex every doc viewURL must match

.env is auto-loaded by config.py (via python-dotenv). You do not need to source it.

Run

Index documents

python scripts/index_documents.py
# optional: wait until Search can find them
python scripts/index_documents.py --verify

Ask from the CLI

python scripts/ask.py "Who owns the Enronry Tony contract?"
python scripts/ask.py "What must OKLightning Tony deliver before go-live?" --top-k 3

MCP (Cursor)

Project config lives at .cursor/mcp.json (absolute path to src/mcp_server.py + env vars). Restart Cursor / reload MCP after editing it. The tool is ask_glean.

Smoke-test the server alone:

python src/mcp_server.py

Tests

python tests/test_pipeline.py

No network; mocks Search/Chat.

How it works

scripts/index_documents.py pushes data/documents.json through the Indexing API. At query time, src/pipeline.py calls Search, then (only if there are hits) Chat with those hits as closed-book context, and returns an answer plus sources. CLI and MCP both call that same pipeline. Results are scoped to our own documents; if Search returns nothing, we suggest close document titles instead of inventing an answer.

Notes / limitations

• Shared sandbox ignores datasourcesFilter; we filter client-side on the document id prefix (CUSTOM_<DATASOURCE>_Contract_).
• Indexing is asynchronous — docs may take a few minutes to become searchable.
• Demo docs use allowAnonymousAccess: true. Do not ship that in production.
• Doc viewURLs must match GLEAN_URL_REGEX or indexing will reject them.

See DESIGN_NOTE.md for API tradeoffs and production notes.