ChemBot

ChemBot is a CLI biomedical Q&A agent built with:

• LangGraph (graph-based agent control loop with multi-route fan-out)
• Model Context Protocol (MCP) (server/client architecture wrapping openFDA and PubMed)
• ChromaDB (persistent local vector store)
• Sentence Transformers (local embeddings)
• OpenAI (LLM responses)

ChemBot answers using ONLY retrieved evidence from three sources:

• FDA Drug Labels — official regulatory prescribing information
• FAERS — FDA Adverse Event Reporting System data
• PubMed — published biomedical literature via NCBI Entrez

---

Architecture

User
  │
  ▼
LangGraph Router (multi-route)
  │
  ├── FDA Label Retrieval ──► openFDA MCP Server
  │
  ├── FAERS Event Retrieval ─► openFDA MCP Server
  │
  └── PubMed Retrieval ─────► PubMed MCP Server
  │
  ▼
Evidence Fusion (groups by source)
  │
  ▼
Answer Generation (OpenAI)

The router can select one or more branches per query:

Question	Routes
"What are the contraindications for ibuprofen?"	FDA Label
"What adverse events have been reported for Ozempic?"	FAERS
"What recent studies exist on semaglutide and Alzheimer's?"	PubMed
"Is ibuprofen safe during pregnancy?"	FDA Label + PubMed
"Is semaglutide associated with pancreatitis?"	FAERS + PubMed

MCP Servers

Both MCP servers use stdio transport — ChemBot automatically spawns and manages them as subprocesses:

[LangGraph Node]
       │ (sync calls)
       ▼
[MCP Client Adapter] (thread-safe async event loop)
       │ (JSON-RPC over stdio)
       ▼
[FastMCP Server]
       │ (requests)
       ▼
[External API]

Server	File	API	Tools
OpenFDA	mcp_servers/openfda_server.py	openFDA	search_drug_labels, search_adverse_events
PubMed	mcp_servers/pubmed_server.py	NCBI Entrez	search_pubmed, search_clinical_trials, search_systematic_reviews, search_meta_analyses, fetch_abstracts

---

Install ChemBot

pip install -r requirements.txt

# Export OpenAI credential
export OPENAI_API_KEY="your-openai-api-key"

# NCBI requires an email for Entrez API usage (PubMed)
export NCBI_EMAIL="your-email@example.com"

Optional API keys

# Higher openFDA rate limits
export OPENFDA_API_KEY="your-openfda-key"

# Higher NCBI rate limits (3 → 10 req/sec)
export NCBI_API_KEY="your-ncbi-key"

---

Run ChemBot

By default, ChemBot runs with all three evidence sources via MCP:

python -m chembot.graph_cli --chroma_dir ./chroma_db

Disable PubMed

python -m chembot.graph_cli --chroma_dir ./chroma_db --disable_pubmed

Use legacy direct OpenFDA client (no MCP)

python -m chembot.graph_cli --chroma_dir ./chroma_db --use_direct_openfda --disable_pubmed

Switch LLM model

python -m chembot.graph_cli --chroma_dir ./chroma_db --openai_model gpt-4-mini

---

Example Questions

FDA Label query

User: What are the contraindications for ibuprofen?

FAERS adverse event query

User: What adverse events have been reported for Ozempic?

PubMed literature query

User: What recent studies exist on semaglutide and Alzheimer's disease?

Multi-source query (FDA Label + PubMed)

User: Is ibuprofen safe during pregnancy?

---

Debug Logging

Enable debug logging to see MCP JSON-RPC traffic:

export CHEMBOT_DEBUG=1
python -m chembot.graph_cli --chroma_dir ./chroma_db

You will see output prefixed with [MCP CLIENT DEBUG] and [PUBMED MCP DEBUG] in stderr:

[MCP CLIENT DEBUG] Calling tool 'search_drug_labels' with arguments: {'search': '...', 'limit': 3}
[MCP CLIENT DEBUG] Parsed tool response: found 3 results. Status: Success
[PUBMED MCP DEBUG] Calling tool 'search_pubmed' with arguments: {'query': '...', 'max_results': 5}
[PUBMED MCP DEBUG] Parsed tool response: found 4 results.

---

Configuration

CLI Arguments

Argument	Default	Description
--chroma_dir	(required)	Path to ChromaDB directory
--collection	chembot_cache	Chroma collection name
--st_model	all-MiniLM-L6-v2	Sentence Transformer model
--k	6	Top-k chunks for answer generation
--label_limit	3	Max FDA label records
--event_limit	3	Max FAERS event records
--pubmed_limit	5	Max PubMed abstracts
--disable_pubmed	false	Disable PubMed retrieval
--use_direct_openfda	false	Use direct HTTP client (no MCP)
--openai_model	gpt-4.1-mini	OpenAI model
--temperature	0.2	LLM temperature
--max_tokens	900	LLM max tokens

Environment Variables

Variable	Required	Description
OPENAI_API_KEY	Yes	OpenAI API key
NCBI_EMAIL	Recommended	Email for NCBI Entrez API (PubMed)
NCBI_API_KEY	No	NCBI API key (increases rate limit)
OPENFDA_API_KEY	No	openFDA API key
CHEMBOT_DEBUG	No	Set to 1 for MCP debug logging
OPENFDA_MCP_COMMAND	No	Custom MCP server command (default: python)
OPENFDA_MCP_ARGS	No	Custom MCP server args (default: mcp_servers/openfda_server.py)
PUBMED_MCP_COMMAND	No	Custom PubMed MCP server command
PUBMED_MCP_ARGS	No	Custom PubMed MCP server args

Custom MCP Server Invocation

By default, clients launch local MCP servers using python mcp_servers/<server>.py. You can customize via environment variables listed above.

For standalone debug mode:

fastmcp dev mcp_servers/openfda_server.py
fastmcp dev mcp_servers/pubmed_server.py