ShinkaEvolve/docs/support_local_models.md

# Local and OpenRouter Models

Shinka supports dynamic LLM backend routing in `LLMClient` and `AsyncLLMClient`.
It also supports dynamic embedding backend routing in `EmbeddingClient` and
`AsyncEmbeddingClient`.
You can use:

- models listed in the provider pricing CSVs (existing behavior)
- dynamic OpenRouter model IDs
- local OpenAI-compatible servers via inline endpoint URIs

## Supported Model Name Formats

### 1) Known models (from `pricing.csv`)

```yaml
evo_config:
  llm_models:
    - gpt-5-mini
    - claude-sonnet-4-6
```

### 2) Dynamic OpenRouter models

Prefix with `openrouter/`:

```yaml
evo_config:
  llm_models:
    - openrouter/qwen/qwen3-coder
    - openrouter/deepseek/deepseek-r1
```

Set env var:

```bash
OPENROUTER_API_KEY=...
```

### 3) Local OpenAI-compatible models

Use `local/<model>@<http(s)://endpoint>`:

```yaml
evo_config:
  llm_models:
    - local/qwen2.5-coder@http://localhost:11434/v1
```

Set optional env var:

```bash
LOCAL_OPENAI_API_KEY=local
```

If not set, Shinka uses `"local"` as a default token.

## Local Embeddings

The same inline local format also works for `embedding_model`.

```yaml
evo_config:
  embedding_model: local/text-embeddings-inference@http://localhost:8080/v1
```

Common local embedding backends:

- Hugging Face TEI:
  `local/text-embeddings-inference@http://localhost:8080/v1`
- vLLM or another OpenAI-compatible embedding server:
  `local/BAAI/bge-small-en-v1.5@http://localhost:8000/v1`
- Ollama OpenAI-compatible endpoint:
  `local/embeddinggemma@http://localhost:11434/v1`

## Notes

- Dynamic OpenRouter/local model IDs are allowed even if not listed in `pricing.csv`.
- If a model has no pricing entry and the provider does not return cost metadata, Shinka records cost as `0.0`.
- Local OpenAI-compatible backend path currently uses chat-completions style calls.
- Local embedding backends use the OpenAI-compatible `/v1/embeddings` path.
- Structured output is not supported yet for `local/...@...` models.

## Applies to Which Clients

These formats work across all LLM consumers that use `LLMClient` / `AsyncLLMClient`, including:

- mutation LLMs (`llm_models`)
- meta LLMs (`meta_llm_models`)
- novelty judge LLMs (`novelty_llm_models`)
- prompt evolution LLMs (`prompt_llm_models`)

For embeddings, the same format applies to:

- code similarity embeddings (`embedding_model`)