Multi-Agent Chatbot with LaunchDarkly AI Configs

Demo application for the Agent Graphs tutorial.

Quick Start

cp .env.example .env
# Add your LD_SDK_KEY and API keys to .env

uv sync
uv run python initialize_embeddings.py --force
uv run python bootstrap/create_configs.py

# Terminal 1
uv run uvicorn api.main:app --reload --port 8000

# Terminal 2
uv run streamlit run ui/chat_interface.py --server.port 8501

Open http://localhost:8501

---

File Structure

Entry Points

File	Purpose
api/main.py	FastAPI server - /chat endpoint
ui/chat_interface.py	Streamlit chat UI

LaunchDarkly Integration

File	Purpose
config_manager.py	Core LD integration - fetches AI Configs and Agent Graphs
api/services/agent_service.py	Graph orchestration - traverses Agent Graph, routes between agents
bootstrap/create_configs.py	Creates AI Configs, tools, segments, and targeting rules in LD

Agents

File	Purpose
agents/supervisor_agent.py	Routes requests based on PII likelihood (uses LD AI Config)
agents/security_agent.py	Detects and redacts PII (uses LD AI Config)
agents/support_agent.py	Answers questions with tools (uses LD AI Config + tools)
agents/ld_agent_helpers.py	Shared helpers - model creation, metric tracking

Tools

File	Purpose
tools_impl/dynamic_tool_factory.py	Creates tools from LD AI Config definitions
tools_impl/search_v1.py	Basic keyword search
tools_impl/search_v2.py	Semantic vector search
tools_impl/reranking.py	BM25 relevance scoring

Data & Utils

File	Purpose
data/vector_store.py	FAISS vector store for search
data/enterprise_kb.py	Knowledge base loader
initialize_embeddings.py	Builds vector index from PDFs
utils/logger.py	Logging helpers

---

How LaunchDarkly Integrates

┌─────────────────────────────────────────────────────────────┐
│                    LaunchDarkly Dashboard                    │
│  • Agent Graph (chatbot-flow)                               │
│  • AI Configs (supervisor, security, support)               │
│  • Targeting rules (free/paid, EU/other)                    │
└─────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────┐
│                     config_manager.py                        │
│  • get_agent_graph() → fetches graph topology               │
│  • get_config() → fetches AI Config for each agent          │
│  • Handles targeting based on user context                  │
└─────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────┐
│                 api/services/agent_service.py                │
│  • Traverses graph by following edges (while loop)          │
│  • Routes based on edge handoff metadata                    │
│  • Tracks metrics (duration, tokens, tool calls)            │
└─────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────┐
│                      agents/*.py                             │
│  • Each agent receives config from LD                       │
│  • Uses config.instructions, config.model                   │
│  • Tracks metrics via config.tracker                        │
└─────────────────────────────────────────────────────────────┘

What You Can Change Without Deploying

Change	Deploy Required?
Model (gpt-4o → claude-sonnet)	No
Prompt/instructions	No
Graph edges (routing)	No
Targeting rules (user segments)	No
Variation rollout percentages	No
New agent type	Yes (register in AGENT_REGISTRY)
New tool integration	Yes (implement in tools_impl/)
Routing contract changes	Yes (update _select_next_node())

Executor Architecture

The AgentGraphExecutor is stateless and generic:

• AGENT_REGISTRY: Maps node key patterns to factory functions
• Edge-following: while current_node loop until terminal node
• Routing: Matches routing_decision to edge handoff.route
• Safety: Cycle detection, max hop limit (10)

To add a new agent type:

# In agent_service.py
AGENT_REGISTRY["my_agent"] = create_my_agent