Two-Collection Qdrant RAG Chat (Ollama)

Simple command-line chat program that:

1. Takes a question as a command-line argument.
2. Retrieves relevant chunks from two Qdrant collections.
3. Sends those chunks to an LLM via Ollama.
4. Prints the answer.

• Python 3.9+
• Ollama running locally (chat + embeddings endpoint)
• At least one model pulled (example: llama3.2)
• Qdrant running with two collections that store payload text

1. Start Ollama:

ollama serve

2. Create .env in the project directory:

OLLAMA_MODEL=llama3.2
OLLAMA_EMBED_MODEL=nomic-embed-text
OLLAMA_CHAT_URL=http://localhost:11434/api/chat
OLLAMA_EMBED_URL=http://localhost:11434/api/embed

QDRANT_URL=http://localhost:6333
QDRANT_API_KEY=
QDRANT_COLLECTION_1=collection_one
QDRANT_COLLECTION_2=collection_two
QDRANT_LIMIT=3
QDRANT_TEXT_FIELDS=text,content,chunk,document,page_content

3. Pull required models on the Ollama server:

ollama pull llama3.2
ollama pull nomic-embed-text

4. Run:

python rag_chat.py "What are the main differences between source A and source B?"

python rag_chat.py "your question" \
  [--env .env] \
  [--collection1 <qdrant-collection>] \
  [--collection2 <qdrant-collection>] \
  [--model llama3.2] \
  [--embed-model nomic-embed-text] \
  [--top-k 3] \
  [--ollama-url http://localhost:11434/api/chat] \
  [--ollama-embed-url http://localhost:11434/api/embed] \
  [--qdrant-url http://localhost:6333] \
  [--qdrant-api-key <key>] \
  [--text-fields text,content,chunk,document,page_content] \
  [--show-context]

• The script generates the question embedding via Ollama, then uses vector search in each Qdrant collection.
• The script checks whether chat/embedding models exist on the configured Ollama server and fails with a clear install command if missing.
• For proxied Ollama endpoints that do not expose management APIs (/api/show or /api/tags), model preflight becomes best-effort and the script continues to normal query execution.
• QDRANT_TEXT_FIELDS controls which payload fields are checked for text context (first matching string is used).
• CLI flags override .env values when both are present.
• Use --show-context to inspect selected chunks from each collection.

Build locally:

docker build -t rag-chat:local .

Run:

docker run --rm \
  --env-file .env \
  rag-chat:local \
  "your question here"

You can also pass flags after the question, for example:

docker run --rm --env-file .env rag-chat:local \
  "your question here" --top-k 5 --show-context

Workflow file: .github/workflows/docker-publish.yml

• Publishes to ghcr.io/<owner>/<repo>.
• Triggers only on pushes to version tags matching v* (example: v1.2.3).
• Uses the pushed Git tag as the Docker tag and also publishes latest.
• Runs only once per release push (no duplicate main + tag build), because branch pushes are not configured as triggers.