ibaz/paparats-embed repository overview

⁠paparats-embed

Fast, zero-config embedding server for Paparats MCP⁠ — code and text embeddings on CPU, no GPU required.

ibaz/paparats-embed bundles llama.cpp⁠'s llama-server behind llama-swap⁠ with two embedding models pre-baked and ready on first request. Drop it in, point Paparats at it, and index code and docs — nothing to download, register, or warm up.

⁠Why

Paparats used to embed with Ollama, then briefly with CC-BY-NC jina models. This image ships two permissively licensed (Apache-2.0) embedders on llama-server, ~5-9x faster than Ollama on CPU:

• Permissive licenses. bge-code-v1 and Qwen3-Embedding-0.6B are both Apache-2.0 — safe for commercial/enterprise use (the previous jina models were CC-BY-NC).
• ~5-9x faster than Ollama on the same hardware. The win is bigger on weak servers, where Ollama's per-request overhead dominates.
• Correct pooling. Both models are decoder-based (--pooling last), baked in and cosine-verified. Wrong pooling silently corrupts the vector space.
• Lazy load + idle unload. Models load on first request and unload after EMBED_TTL seconds idle. A code-only user never spins up the text model; RAM is freed when nobody's searching.
• Zero-config, CPU-only. No CUDA, no model registry, no Modelfile. Both GGUFs are baked in.

⁠Models

Both are served on a single OpenAI-compatible endpoint; route by model name.

⁠Run

docker run -d --name paparats-embed \
  -p 11434:8080 \
  -e EMBED_TTL=300 \
  ibaz/paparats-embed:latest

Embed code via the OpenAI-style API:

curl http://localhost:11434/v1/embeddings \
  -H 'Content-Type: application/json' \
  -d '{"model":"bge-code-v1","input":"def add(a, b): return a + b"}'

Embed arch/docs prose by routing to the text model:

curl http://localhost:11434/v1/embeddings \
  -H 'Content-Type: application/json' \
  -d '{"model":"qwen3-embedding-0.6b","input":"How the indexer maps commits to chunks"}'

⁠Configuration

Cold start (first request after idle/unload) is a few seconds while the model loads; every request after that is hot until EMBED_TTL elapses.

⁠Ports

llama-swap listens on 8080 inside the container. The examples map it to host 11434 for a painless migration from Ollama (same host port), but any host port works.

Built for Paparats MCP⁠ — semantic code search over your repositories.

embeddings, embedding-server, llama-cpp, llama-server, llama-swap, mcp, semantic-search, code-search, vector-search, bge-code-v1, qwen3-embedding, bge, gguf, cpu, ollama-alternative, qdrant, apache-2.0