RAG Agent with Qdrant & Ollama

This project implements a simple RAG (Retrieval‑Augmented Generation) agent that uses:

• Qdrant – a vector database for storing embeddings.
• Ollama – local LLM and embedding model (llama3 and nomic‑embed‑text).
• LangChain – framework for building the agent and tools.

The agent can:

• Add documents to the knowledge base.
• Search the knowledge base for relevant chunks.
• Answer user queries using the stored knowledge.

Project structure

workspace/
├── src/
│   ├── __init__.py
│   ├── vector_store.py   # Qdrant wrapper
│   ├── tools.py          # LangChain tools
│   ├── agent.py          # Agent definition
│   ├── loader.py         # Load all .txt files from a directory
│   └── cli.py            # Interactive command‑line client
├── requirements.txt
└── README.md

Installation

# Pull the required Ollama models
ollama pull llama3
ollama pull nomic-embed-text

# Install Python dependencies
pip install -r requirements.txt

Usage

1. Load documents into the knowledge base

python -m src.loader /path/to/text/files

All .txt files in the directory (recursively) are added to the vector store.

2. Start the interactive CLI

python -m src.cli

Once started you can use the following commands:

Command	Description
/add <title> <file_path>	Add a single file to the knowledge base.
/search <query>	Search the knowledge base and display top results.
/quit	Exit the program.
/help	Show help.
Any other text	Sent to the agent as a user query.

3. Example session

RAG Agent CLI. Type /help for commands.
> /add example docs/example.txt
Document 'example' added.
> /search quantum
Results:
1. [example - chunk 0] Quantum mechanics is the branch of physics that deals with...
> Tell me more about quantum.
Sure! Here is what I found in the knowledge base: ...
> /quit
Goodbye.

How it works

• Vector Store – KnowledgeBase wraps a QdrantVectorStore. It creates the collection only if it does not exist, preventing accidental data loss.
• Chunking – Documents are split into 500‑character chunks with 50‑character overlap using RecursiveCharacterTextSplitter.
• Tools – Two LangChain tools are exposed:
  • search_knowledge_base(query, max_results) – returns a list of relevant chunks.
  • add_to_knowledge_base(content, title) – adds a document.
• Agent – Built with create_agent from langchain.agents. It uses the local ChatOllama model (llama3).

Extending

• Replace the embedding model by editing KnowledgeBase.__init__.
• Add more tools (e.g., delete from knowledge base) following the same pattern.
• Deploy the agent as a web service by wrapping run_query in a FastAPI endpoint.

License

MIT License.