aditya-karande/chatbot_with_langgraph

Purpose and Entry Point

This is a Streamlit-based agentic chatbot that combines LangGraph orchestration, LLM tool calling, document RAG, approval-gated code execution, and two memory scopes. Runtime starts at app.py, which hosts chat, thread interaction, file upload, and human approval/resume controls.

Application Boundary and State

The backend models a stateful LangGraph conversation rather than a request/response web service. backend/state.py defines chatbot state and prompting context; backend/config.py centralizes model, embedding, database, and memory/checkpoint setup. Conversation identity and user context must flow consistently through graph invocations and resumes.

Main Orchestration Graph

backend/graph.py composes the primary LangGraph workflow and routes among agent generation, tools, RAG, memory persistence, and summarization. backend/agent.py supplies the main agent decision logic and conditional paths. The graph is the control-plane boundary: individual capabilities are graph nodes or subgraphs, while Streamlit invokes/resumes the compiled workflow.

Agent and Tool Execution

The agent uses an external Groq-hosted openai/gpt-oss-120b model to decide whether to answer directly or request capabilities. backend/tools.py exposes weather, stock, currency, date, web search, and Python execution tools. Safe tool calls return into the agent loop; Python execution crosses a human-approval boundary through hitl_gate in backend/agent.py, using LangGraph interruption until the UI approves or rejects.

Retrieval-Augmented Generation

backend/rag.py owns document upload/ingestion and a dedicated LangGraph RAG subgraph. PDF, TXT, and Markdown content is chunked, embedded with Gemini embeddings, and persisted in PostgreSQL pgvector. At query time, retrieval is followed by relevance grading/filtering before selected chunks are supplied for answer generation. This separates document knowledge from general agent/tool reasoning.

Memory and Conversation Lifecycle

Short-term conversation continuity is checkpoint-backed and managed with history compression in backend/summarization.py; summaries replace older messages when context grows. Long-term, cross-conversation facts are retrieved before response generation and selectively deduplicated/stored afterward by backend/memory.py. backend/threads.py manages user-facing chat histories, titles, and renaming, distinct from graph checkpoint state.

Runtime Infrastructure

The application runs as Python/Streamlit (app.py) with LangGraph as the workflow runtime. It depends on Groq for LLM inference, Google/Gemini for embeddings, PostgreSQL plus pgvector for vector persistence, and external APIs for tools such as weather and search. Configuration and connection initialization are concentrated in backend/config.py; deployment requires database connectivity, pgvector enabled, and API credentials.