cadrumo.adapters.outbound.llm.providers.local module

Local provider adapter for Ollama-compatible runtimes.

Speaks the Ollama /api/chat endpoint (resolved per call from Settings.cadrumo_llm_ollama_chat_url) and adapts its response into the ProviderCompletion shape. The adapter assumes the runtime is reachable on localhost; remote Ollama deployments are out of scope.

For multimodal evidence, rasterise_pdf_pages_to_base64_png() renders an in-memory PDF to base64 PNG pages fully on-host so a local vision model can read a scan-only invoice; the adapter forwards those base64 images on the Ollama images message field. No file is written and nothing leaves the host (sensitive-financial-data-secure-storage-only).

rasterise_pdf_pages_to_base64_png(pdf_bytes, *, scale=2.0)[source]

Rasterise each page of an in-memory PDF to a base64-encoded PNG, on-host.

Renders every page in process memory via pypdfium2 and Pillow so a local vision model can read a scan-only or image-only PDF that has no extractable text layer. Nothing is written to disk and nothing leaves the host (sensitive-financial-data-secure-storage-only).

Parameters:

  • pdf_bytes (bytes) – In-memory PDF bytes (read transiently from secure storage).

  • scale (float) – pypdfium2 render scale; a larger value yields a larger raster.

Return type:

tuple[str, ...]

Returns:

One base64-encoded PNG string per page, in page order.

Raises:
class LocalAdapter(timeout_s)[source]

Bases: ProviderAdapter

Provider adapter that invokes a local Ollama-compatible HTTP endpoint.

Parameters:

timeout_s (int)

provider: LLMProvider
supports_images: bool
async complete(request)[source]

Execute a chat completion request against the local endpoint.

Parameters:

request (ProviderRequest) – Normalized provider request.

Return type:

ProviderCompletion

Returns:

ProviderCompletion containing the trimmed assistant message and reported token counts.

Raises:

LLMProviderError – When the runtime returns a non-2xx HTTP error status.