7f7223e0ac
Replaces gpt-oss on ModelConfig, qa.model and processing.title_model, and ministral-3 on the picture-description model. qa.model.vision follows the model and is now true. enable_thinking was gated on the gpt-oss name, so it did nothing for qwen3.8. With title_model's max_tokens of 100 the reasoning consumed the whole budget and title generation returned an empty string. The mapping now applies to any ollama model via reasoning_effort(): false sends "none", true sends "high". Measured on qwen3.8:27b-mlx, "low" does not disable thinking and "none" does; gpt-oss is the inverse, its template has no "none" level, so it keeps "low". Picture description bypasses get_model -- docling posts the request itself from a params dict -- so the flag was inert on that path too. vlm_api_params() carries reasoning_effort into both converters' request bodies. At max_tokens 200 the description survived either way, but the switch cut completion tokens from 141 to 45. test_search_tool_skips_binary_content_when_qa_model_is_text_only asserted the vision default rather than setting it; it now configures vision=False itself. docs/benchmarks.md keeps ministral-3: those are recorded measurements.
124 lines
4.1 KiB
Python
124 lines
4.1 KiB
Python
"""Base class for document converters."""
|
|
|
|
import os
|
|
from abc import ABC, abstractmethod
|
|
from pathlib import Path
|
|
from typing import TYPE_CHECKING
|
|
|
|
if TYPE_CHECKING:
|
|
from docling_core.types.doc.document import DoclingDocument
|
|
|
|
from haiku.rag.config import AppConfig
|
|
from haiku.rag.config.models import ModelConfig
|
|
|
|
|
|
def vlm_api_url(config: "AppConfig", model: "ModelConfig") -> str:
|
|
"""Construct the VLM chat-completions URL for a picture-description model."""
|
|
if model.base_url:
|
|
return f"{model.base_url.rstrip('/')}/v1/chat/completions"
|
|
|
|
if model.provider == "ollama":
|
|
return f"{config.providers.ollama.base_url.rstrip('/')}/v1/chat/completions"
|
|
|
|
if model.provider == "openai":
|
|
return "https://api.openai.com/v1/chat/completions"
|
|
|
|
raise ValueError(f"Unsupported VLM provider: {model.provider}")
|
|
|
|
|
|
def vlm_api_headers(model: "ModelConfig") -> dict[str, str]:
|
|
"""Auth headers for the picture-description VLM endpoint. Docling posts to
|
|
it directly, so the key travels as a header rather than through an SDK.
|
|
|
|
The public OpenAI endpoint falls back to ``OPENAI_API_KEY``. A custom
|
|
``base_url`` never does: that key belongs to api.openai.com, not to
|
|
whatever self-hosted server the model points at.
|
|
"""
|
|
key = model.api_key
|
|
if not key and model.provider == "openai" and not model.base_url:
|
|
key = os.environ.get("OPENAI_API_KEY")
|
|
if key:
|
|
return {"Authorization": f"Bearer {key}"}
|
|
return {}
|
|
|
|
|
|
def vlm_api_params(model: "ModelConfig", max_tokens: int) -> dict[str, object]:
|
|
"""Request body fields docling posts alongside the picture."""
|
|
from haiku.rag.utils import reasoning_effort
|
|
|
|
params: dict[str, object] = {
|
|
"model": model.name,
|
|
"max_completion_tokens": max_tokens,
|
|
}
|
|
effort = reasoning_effort(model)
|
|
if effort is not None:
|
|
params["reasoning_effort"] = effort
|
|
return params
|
|
|
|
|
|
class DocumentConverter(ABC):
|
|
"""Abstract base class for document converters.
|
|
|
|
Document converters are responsible for converting various document formats
|
|
(PDF, DOCX, HTML, etc.) into DoclingDocument format for further processing.
|
|
"""
|
|
|
|
@property
|
|
@abstractmethod
|
|
def supported_extensions(self) -> list[str]:
|
|
"""Return list of file extensions supported by this converter.
|
|
|
|
Returns:
|
|
List of file extensions (including the dot, e.g., [".pdf", ".docx"]).
|
|
"""
|
|
pass
|
|
|
|
@abstractmethod
|
|
async def convert_file(
|
|
self, path: Path, source_uri: str | None = None
|
|
) -> "DoclingDocument":
|
|
"""Convert a file to DoclingDocument format.
|
|
|
|
Args:
|
|
path: Path to the file to convert.
|
|
source_uri: Optional origin URI (e.g. the URL the file was
|
|
downloaded from) used by docling's HTML/Markdown backends to
|
|
resolve relative `<img src="/path">` references. Ignored by
|
|
converters that have no equivalent backend option (notably
|
|
docling-serve).
|
|
|
|
Returns:
|
|
DoclingDocument representation of the file.
|
|
|
|
Raises:
|
|
ValueError: If the file cannot be converted.
|
|
"""
|
|
pass
|
|
|
|
SUPPORTED_FORMATS = ("md", "html", "plain")
|
|
|
|
@abstractmethod
|
|
async def convert_text(
|
|
self,
|
|
text: str,
|
|
name: str = "content.md",
|
|
format: str = "md",
|
|
source_uri: str | None = None,
|
|
) -> "DoclingDocument":
|
|
"""Convert text content to DoclingDocument format.
|
|
|
|
Args:
|
|
text: The text content to convert.
|
|
name: The name to use for the document (defaults to "content.md").
|
|
format: The format of the text content ("md", "html", or "plain").
|
|
Defaults to "md". Use "plain" for plain text without parsing.
|
|
source_uri: Optional origin URI used by docling's HTML/Markdown
|
|
backends to resolve relative image references.
|
|
|
|
Returns:
|
|
DoclingDocument representation of the text.
|
|
|
|
Raises:
|
|
ValueError: If the text cannot be converted or format is unsupported.
|
|
"""
|
|
pass
|