Code-QA

This project is a lightweight code-aware agent that can answer questions about a repository.

It enables natural language querying over a local codebase and retrieves relevant code snippets to support answers.**

The end goal is to add functionality till an orchestration pattern is required. I want to know the most an AI agent can do within the bounds of good design practises.

---

Overview

RAG-powered Agentic AI system that combines:

• Semantic code retrieval
• LLM reasoning loop
• Tool-based execution

to create an interactive AI assistant for repositories.

---

Retrieval Layer (RAG)

Code is embedded into a vector store using FAISS.
Currently, the codebase must be manually indexed before running the agent.

Workflow:

• Code is converted into embeddings and stored
• A retriever fetches relevant chunks based on semantic similarity
• The LLM uses retrieved context to generate grounded answers

---

Agent (ReAct Pattern)

The system uses an agent loop based on the ReAct pattern:

Thought → Action → Action Input → Observation → Final Answer

Available Tool:

• search_codebase(query) → retrieves relevant code snippets

Agent Behavior:

The agent dynamically decides:

• when to search the codebase
• what query to use

---

🧩 Design Pattern / Architecture

Layered Architecture

User Query
↓
Agent (ReAct loop)
↓
Tool Layer (search_codebase)
↓
Retriever (Vector DB)
↓
Embedded Codebase (FAISS)

---

System Workflow

1. User asks a question
2. Agent analyzes the query (Thought step)
3. Agent calls search_codebase
4. Retriever fetches relevant code chunks
5. Agent receives observations
6. Agent generates final answer

---

Tech Stack

• LangChain
• Google Gemini (ChatGoogleGenerativeAI)
• FAISS (Vector Store for embeddings)
• Python unittest (testing)
• LangSmith (prompt management)

---

Summary

This project is a modular AI system combining:

• Retrieval (RAG)
• Reasoning (Agent loop)
• Execution (Tools)

to build a practical AI-powered code assistant.

---

Future Improvements

• Switch to tool-calling agent (more stable than ReAct)
• Better chunking strategy, say, AST (Abstract Syntax Tree), maybe a code graph
• Enable hybrid retrieval combining semantic search (embeddings) and keyword search (function names, variables)
• Add some kind of memory to allow the agent to remember stuff.
• Add support for querying remote repositories (e.g., GitHub)
• Evaluation - define metrics
• Packaging
• CI/CD
• Add support for more file formats. Currently only .py is supported. Single-loader to a multi-loader approach