Chem Research

Live chat: https://chem-research-web.vercel.app · v0.1 shipped (literature search & citation)

MCP server + literature RAG over open combustion / QSAR / fuel-chemistry papers. Ask in the browser or wire Cursor / Claude Desktop to the same tools.

Demo: chat → answer → MCP connect → evals

Try it

Path	URL
Chat	chem-research-web.vercel.app
Evals	/evals
MCP connect (demo key + copy-paste)	/mcp

No account for chat (Supabase anonymous session + daily quota). Cold starts are intentional: generate has no keep_warm (spend control); first answers after idle take ~1–2 minutes.

Connect over MCP

Endpoint (trailing slash required):
https://tjkessler--chem-research-api-fastapi-app.modal.run/mcp/

Demo API key: copy from the live /mcp page (low-quota public demo, 25/day). Prefer env substitution — do not commit the plaintext key.

Cursor (.cursor/mcp.json or Settings → MCP):

{
  "mcpServers": {
    "chem-research": {
      "url": "https://tjkessler--chem-research-api-fastapi-app.modal.run/mcp/",
      "headers": {
        "X-API-Key": "${env:MCP_DEMO_API_KEY}"
      }
    }
  }
}

Set MCP_DEMO_API_KEY to the demo key from /mcp. Full Claude Desktop / mcp-remote notes: docs/MCP_CLIENTS.md.

Tools: search_literature, fetch_citation.

Architecture

flowchart TB
    subgraph Clients
        A["MCP client<br/>(Claude Desktop / Code)"]
        B["Web app<br/>(Next.js on Vercel)"]
    end

    subgraph Backend["Backend: FastAPI + MCP — Modal ASGI app"]
        C["/mcp — MCP tools<br/>(streamable HTTP)"]
        D["/api — REST<br/>(chat, admin, evals)"]
        E["core/ — tools, retrieval,<br/>ingestion, cost, rate limits"]
    end

    subgraph ModalGPU["Modal: LLM inference"]
        F["embed()<br/>bge-small-en-v1.5, 384-d"]
        G["generate()<br/>Qwen2.5-7B-Instruct · L4"]
    end

    subgraph SB["Supabase"]
        H[("Postgres + pgvector")]
        I["Auth<br/>admin + anon quotas"]
    end

    A -->|streamable HTTP + X-API-Key| C
    B -->|HTTPS| D
    C --> E
    D --> E
    E --> F
    E --> G
    E --> H
    D --> I
    C -.->|API key → quota| I

Piece	Host
Frontend	Vercel Hobby — apps/web
API + MCP	Modal ASGI — chem-research-api
Embed / generate	Modal — chem-research-embed, chem-research-generate
Postgres + Auth	Supabase

Eval ablation (chunking × retrieval)

Retrieval-only on evals/dataset.jsonl (n=38), top_k=8, P@5 / R@5. Directional only — small labeled set. Full write-up: evals/results/ABLATION.md.

Chunking	Mode	Mean P@5	Mean R@5	Mean latency (ms)
tuned	vector	0.195	0.759	959
tuned	hybrid	0.200	0.785	680
naive	vector	0.216	0.829	647
naive	hybrid	0.216	0.829	868

Default: tuned chunks (1000/150) + hybrid retrieval. Naive wins raw doc-level P/R but ~1.8× more chunks (confound); see ablation notes.

Live dashboard: chem-research-web.vercel.app/evals.

Cost & latency (measured)

Production measurements (2026-08-06, Modal L4 generate, keep_warm=0):

Scenario	Latency	Cost (est.)	Source
Public chat, GPU cold start (“What is cetane number?”)	95–108 s	~$0.022–0.024	Live Vercel → Modal /api/chat
MCP search_literature e2e (cold generate)	~99 s	same L4 wall accounting	verify_mcp_remote
Retrieval-only eval mean (ablation, n=38)	~0.6–1.0 s	~$0.0002 / query	[evals/results/ABLATION.md](evals/results/ABLATION.md)

Cold-start policy: no keep_warm on generate — honest UX instead of burning the free-plan credit. scaledown_window=180s. Chat shows a waking hint after 4s while a request is in flight. Warm hits (container still up) are much faster; we cite cold numbers because users often hit an idle demo.

Cost model: soft token estimate (chars/4) + Modal L4 wall time at $0.000222/s (raw latency_ms/1000, no floor).

Operational Modal target: ≤$25/mo under the free-plan $30 credit (concurrency caps; no dashboard hard budget on free).

Local setup

uv sync --group dev
cp .env.example .env
# fill Supabase + optional Modal / MCP demo key — see docs/SETUP.md
uv run uvicorn apps.api.main:app --reload --port 8000

Frontend: cd apps/web && npm install && npm run dev (see apps/web/README.md).

More: docs/SETUP.md.