Chat Format

Format prompts for multi-provider LLM inference with context retrieval.

When to use

When you need a prompt string in a local model's chat template (for llama.cpp, Ollama or another local runtime), or a small embedding-based router over a handful of patterns. It does not call a model and it does not format for hosted APIs: Anthropic, OpenAI and Gemini take structured messages, not a template string.

Steps

1. Format chat — call mcp__plugin_ruflo-core_ruflo__ruvllm_chat_format with messages ([{role, content}]) and template: one of llama3, mistral, chatml, phi, gemma, or a model id to auto-detect. Both fields are required. It returns the formatted string. It needs @ruvector/ruvllm-wasm (npm i -g @ruvector/ruvllm-wasm); without it the call fails with an install hint.
2. Create a router (optional) — ruvllm_hnsw_create with dimensions (for example 384), maxPatterns, efSearch. The WASM router holds about 11 patterns at most (ruvllm-wasm v2.0.1), so use it for routing between a few named intents, not for document retrieval. For real document search use memory_search or the ruflo-rag-memory plugin.
3. Add patterns — ruvllm_hnsw_add with routerId, a name, and an embedding whose length equals the router's dimensions (you produce the embedding, for example with embeddings_generate).
4. Route a query — ruvllm_hnsw_route with routerId, the query embedding, and k (default 3).
5. Check status — ruvllm_status shows whether the WASM module loaded. Routers live in the MCP server process only; they are gone when it restarts.