Skip to content

README.md

Latest commit

 

History

History
52 lines (35 loc) · 3.71 KB

README.md

File metadata and controls

52 lines (35 loc) · 3.71 KB
Raw

Browser demo (Data360 MCP agent)

Minimal FastAPI server + static page that streams LangGraph v2 events (tokens + tools) via Server-Sent Events.

Prerequisites

Same as other agent examples: running Data360 MCP server, DATA360_MCP_URL, OPENAI_API_KEY, and this repo synced with uv sync so data360-mcp-agent is available.

You can put those (and optional DATA360_AGENT_MODEL, etc.) in a .env file:

  • data360-mcp/.env at the repository root — loaded automatically when you start uvicorn from anywhere under this repo.
  • .env in your current working directory — loaded second; duplicate keys override the repo file (useful when you run the server from another folder).

Shell exports still win over .env unless you clear them.

Run

From the repository root (data360-mcp/):

uv run uvicorn demo_web.server:app --app-dir examples/agents/langchain-graph --host 127.0.0.1 --port 8844

Or set variables explicitly instead of .env:

export DATA360_MCP_URL=http://127.0.0.1:8000/mcp
export OPENAI_API_KEY=...
uv run uvicorn demo_web.server:app --app-dir examples/agents/langchain-graph --host 127.0.0.1 --port 8844

Open http://127.0.0.1:8844/ in a browser, enter a query, click Stream.

The Answer panel shows the assistant’s final narrative: the last on_chat_model_end event whose output has no tool calls (typical closing synthesis after tools). If the graph ends differently, the UI falls back to the last chat-model message text.

Notes

  • Local dev only — no authentication; bind to 127.0.0.1 as above.
  • Health: GET /api/health
  • Stream: POST /api/stream with JSON {"query": "..."} — response is text/event-stream.
  • K360 staged query: POST /api/k360-query with JSON {"query":"..."} — returns {gate, rewrite, tool_calls, content_packet, narrative, error}.
  • Viz charts: When the agent calls data360_get_viz_spec / data360_get_multi_indicator_viz_spec, the page listens for LangGraph on_tool_end events, reads the url in the tool result (static …/static/viz_specs/*_vega.json), and fetches the Vega-Lite JSON via GET /api/fetch-viz-spec?url=… (server-side proxy avoids browser CORS). The chart is rendered with vega-embed from a CDN — the same spec as @data360/mcp-ui VegaChartCard, without bundling React.
  • WEBSITE_HOSTNAME: The MCP server builds chart URLs with WEBSITE_HOSTNAME or defaults to http://localhost:{MCP_PORT}. Set WEBSITE_HOSTNAME to the origin your browser can reach (e.g. http://127.0.0.1:8000) so it matches DATA360_MCP_URL’s host/port; otherwise the proxy may reject cross-host URLs (localhost vs 127.0.0.1 is allowed when ports match).
  • The compiled agent is cached after the first successful request.
  • POST /api/k360-query uses the staged K360 pipeline (Gate -> Rewriter -> Compile -> Narrative).

Troubleshooting (chart not showing)

  • Tool output was not JSON in the stream: LangGraph’s v2 events used to carry ToolMessage objects, which json.dumps turned into useless repr strings, so the page could not read url. The demo server now normalizes on_tool_end → data.output into plain dicts before SSE.
  • Query must actually call a viz tool (e.g. ask for a chart, or a flow that runs data360_get_viz_spec after fetching data). A text-only answer will not produce a spec URL.
  • Proxy 400: GET /api/fetch-viz-spec only allows the static …/static/viz_specs/*_vega.json path on the same port as DATA360_MCP_URL and a safe host. Set WEBSITE_HOSTNAME on the MCP process to match the host you use in the browser (see .env.example).