| from __future__ import annotations |
|
|
| from app.config import Settings |
| from app.route import resolve_brain |
| from backends.apple import AppleEmbed, AppleLLM, AppleOCR |
| from backends.base import EmbedBackend, LLMBackend, OCRBackend |
| from backends.cpu import CpuOCR |
| from backends.gemma import GemmaEmbed, GemmaLLM |
| from backends.nvidia import NvidiaEmbed, NvidiaLLM, NvidiaOCR |
| from backends.ollama import OllamaEmbed, OllamaLLM, OllamaOCR |
|
|
| __all__ = [ |
| "EmbedBackend", |
| "LLMBackend", |
| "OCRBackend", |
| "build_embed", |
| "build_llm", |
| "build_ocr", |
| ] |
|
|
|
|
| def _with_brain(settings: Settings) -> Settings: |
| """Point llm/embed URLs at the resolved Gemma 4 12B (direct vLLM or Hermes).""" |
| route = settings.llm_route.lower().strip() |
| backend = settings.llm_backend.lower().strip() |
| if route == "studio": |
| return settings |
| if backend in {"ollama", "lightning", "apple"}: |
| return settings |
| try: |
| brain = resolve_brain(settings) |
| except RuntimeError: |
| return settings |
| if brain.kind == "studio": |
| return settings |
| updates: dict[str, object] = {} |
| if brain.base_url and brain.base_url != settings.llm_base_url: |
| updates["llm_base_url"] = brain.base_url |
| updates["embed_base_url"] = brain.base_url |
| if brain.model and brain.model != settings.llm_model: |
| updates["llm_model"] = brain.model |
| if settings.embed_backend.lower() in {"omni", "gemma", "gemma4"}: |
| updates["embed_model"] = brain.model |
| return settings.model_copy(update=updates) if updates else settings |
|
|
|
|
| def build_llm(settings: Settings, *, client=None) -> LLMBackend: |
| settings = _with_brain(settings) |
| name = settings.llm_backend.lower().strip() |
| if name in {"gemma", "gemma4", "unified", "hermes", "auto"}: |
| return GemmaLLM(settings, client=client) |
| if name in {"nvidia", "vllm", "qwen", "qwen38"}: |
| return NvidiaLLM(settings, client=client) |
| if name in {"ollama", "lightning"}: |
| return OllamaLLM(settings, client=client) |
| if name == "apple": |
| return AppleLLM() |
| raise ValueError(f"unknown llm_backend: {settings.llm_backend}") |
|
|
|
|
| def build_embed(settings: Settings, *, client=None) -> EmbedBackend: |
| settings = _with_brain(settings) |
| name = settings.embed_backend.lower().strip() |
| if name in {"omni", "gemma", "gemma4"}: |
| return GemmaEmbed(settings, client=client) |
| if name in {"nvidia", "nemotron", "vllm"}: |
| return NvidiaEmbed(settings, client=client) |
| if name in {"openai", "ollama"}: |
| return OllamaEmbed(settings, client=client) |
| if name == "apple": |
| return AppleEmbed() |
| raise ValueError(f"unknown embed_backend: {settings.embed_backend}") |
|
|
|
|
| def build_ocr(settings: Settings, *, client=None) -> OCRBackend | None: |
| name = settings.ocr_backend.lower().strip() |
| if name in {"", "none", "off"}: |
| return None |
| if name == "ollama": |
| return OllamaOCR(settings, client=client) |
| if name == "nvidia": |
| return NvidiaOCR() |
| if name == "apple": |
| return AppleOCR() |
| if name == "cpu": |
| return CpuOCR() |
| raise ValueError(f"unknown ocr_backend: {settings.ocr_backend}") |
|
|
