Insights Generator (LangGraph Multi-Agent)

Multi-agent analytics pipeline for CSV/Parquet data with configurable LLM backend.

Features

• Upload CSV/Parquet data.
• Intent recognition with clarification loop.
• Data analytics agent computes:
  • mean, median, mode, average
  • anomaly detection using IQR
  • long-tail detection (skew)
  • high-variance detection (coefficient of variation)
• Visualization agent generates Plotly charts from templates.
• Insight agent writes trend/findings summary.
• Swappable model backend via environment variables.

Model swapping

Set in .env:

• MODEL_PROVIDER: openai, anthropic, or none
• MODEL_NAME: model id for the provider
• PROMPTS_PATH: YAML prompt pack path (default prompts/insights_prompts.yaml)

If provider config is missing/unavailable, the app falls back to deterministic heuristics.

Externalized prompts and few-shots

• Prompt pack file: prompts/insights_prompts.yaml
• intent section contains parser rules and intent few-shot JSON examples.
• insight section contains business logic constraints and insight few-shot examples.
• Update this YAML to tune domain logic without code changes.

Setup

python -m venv .venv
source .venv/bin/activate
pip install -e .
pip install -e .[providers]
uvicorn insights_generator.api:app --reload

Open docs: http://127.0.0.1:8000/docs

Endpoints

• POST /analyze (multipart form)
  • file: CSV or Parquet
  • user_prompt: optional
  • use_python_repl: optional bool
  • use_mcp: optional bool
• POST /clarify (JSON)
  • session_id
  • clarification

Generated charts are saved in artifacts/<session_id>/.

Agentic chatbot example (clarify + joke + recipe)

This repo also includes a maintainable/testable agentic chatbot module in src/agentic_chatbot:

• planner.py: intent routing planner (clarify, joke, recipe)
• skills.py: isolated skills (clarification, joke generation, recipe generation)
• agent.py: orchestration layer for dispatching plans to skills
• llm.py: OpenAI adapter behind an LLMClient protocol for dependency injection
• factory.py: ChatbotFactory for provider swapping (openai, anthropic, google)
• mcp.py: MCP connectors/registry for external tool and prompt servers

Why this is testable

• Planner, agent, and skills use interface-based dependency injection.
• Unit tests use stubs/mocks, so no real OpenAI calls are needed.
• OpenAI payload formatting is verified in isolation.

Run tests

pip install -e .[dev]
pytest

Run chatbot CLI

export OPENAI_API_KEY=your_key
export OPENAI_MODEL=gpt-4o-mini
export LLM_PROVIDER=openai
agentic-chatbot

Provider-specific env keys:

• OpenAI: LLM_PROVIDER=openai, OPENAI_API_KEY, OPENAI_MODEL
• Anthropic: LLM_PROVIDER=anthropic, ANTHROPIC_API_KEY, ANTHROPIC_MODEL
• Google: LLM_PROVIDER=google, GOOGLE_API_KEY, GOOGLE_MODEL

MCP connectors (tools + prompts)

Use MCPConnectorRegistry to register aliases for MCP prompt and tool endpoints, then inject it into ChatbotFactory.

from agentic_chatbot.factory import ChatbotFactory
from agentic_chatbot.mcp import MCPConnectorRegistry, MCPPromptConnector, MCPToolConnector

registry = MCPConnectorRegistry()
registry.register_prompt(
    "joke_system",
    MCPPromptConnector(client=mcp_client, server="prompt-server", prompt_name="joke_prompt"),
)
registry.register_tool(
    "joke_policy",
    MCPToolConnector(client=mcp_client, server="tools-server", tool_name="policy_lookup"),
)

agent = ChatbotFactory(provider="openai", mcp_registry=registry).build_agent()

Planner params can reference connectors:

• mcp_prompt: prompt alias (for system prompt override)
• mcp_tool: tool alias (tool output is appended as extra context)
• prompt_args / tool_args: optional argument dictionaries