ibaz/paparats-ollama repository overview

⁠paparats-ollama

Minimal Ollama with pre-baked Jina Code Embeddings 1.5B⁠ — model ready to serve on container start.

Part of the Paparats MCP⁠ semantic code search platform. This image eliminates the need to install Ollama on your host or download models manually. Just docker run and start generating code embeddings.

⁠Why Use This Image?

• Zero setup — no Ollama binary install, no GGUF model download, no Modelfile configuration
• Instant startup — model is pre-registered and ready to serve immediately
• Lightweight — built on alpine/ollama (~70 MB base), total ~1.5 GB including the model
• CPU-only — no GPU required, works on any machine (sufficient for embedding generation)
• Consistent — same model version across all environments (dev, CI, production)

⁠Quick Start

docker run -d -p 11434:11434 ibaz/paparats-ollama

Verify the model is ready:

curl http://localhost:11434/api/tags
# Should list "jina-code-embeddings" model

Generate embeddings:

curl http://localhost:11434/api/embed -d '{
  "model": "jina-code-embeddings",
  "input": "function authenticate(user, password) { ... }"
}'

⁠Model Details

Supports task-specific prefixes for better relevance:

• nl2code — natural language to code search
• code2code — find similar code patterns
• techqa — technical question answering

⁠Use with Paparats MCP

This image is designed to work with the Paparats semantic code search stack:

# Install the full stack with Docker Ollama (no host Ollama needed)
npm install -g @paparats/cli
paparats install --ollama-mode docker

Or use in a Docker Compose setup:

services:
  ollama:
    image: ibaz/paparats-ollama
    ports:
      - "11434:11434"

  paparats:
    image: ibaz/paparats-server
    environment:
      OLLAMA_URL: http://ollama:11434
    ports:
      - "9876:9876"

  qdrant:
    image: qdrant/qdrant
    ports:
      - "6333:6333"

⁠Use Standalone

This image works as a general-purpose code embedding server, independent of Paparats:

import requests

response = requests.post("http://localhost:11434/api/embed", json={
    "model": "jina-code-embeddings",
    "input": "async function fetchUser(id) { return await db.users.findById(id); }"
})
embedding = response.json()["embeddings"][0]  # 1536-dimensional vector

Use it for:

• Semantic code search engines
• Code similarity detection
• RAG pipelines over codebases
• Code clustering and analysis

⁠How It's Built

Multi-stage Docker build for minimal image size:

1. Stage 1 — Uses official ollama/ollama to download the GGUF model from HuggingFace and register it via ollama create
2. Stage 2 — Copies the registered model into lightweight alpine/ollama base (~70 MB)

Result: ~1.5 GB image (vs ~5+ GB if using the full Ollama image with runtime download).

⁠Tags

• latest — most recent stable build
• x.y.z — version-pinned releases (e.g., 0.2.0)

⁠Health Check

Built-in health check polls /api/tags every 10 seconds:

docker inspect --format='{{.State.Health.Status}}' <container_id>

⁠Links

• Paparats MCP⁠ — semantic code search for AI coding assistants
• Jina Code Embeddings⁠ — the embedding model
• Ollama⁠ — local LLM runtime

⁠License

MIT — github.com/bkbhub/paparats-mcp⁠