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

Related MCP server: search-context

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

.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.