Skip to content

Toolkit Modules Coverage

This page documents how source modules are included in the documentation set.

Inclusion strategy

The docs use a mixed strategy:

  • Full API reference for stable, high-usage modules
  • Brief conceptual overview for configuration-heavy or data/prompt-heavy modules

This balance keeps human docs readable while preserving deep API access where it matters.

Requested modules and coverage

src/doc_processing_toolkit/data/model_information.json

  • Coverage type: brief overview
  • Overview: Canonical model registry used across the package for provider routing, token/cost pricing lookups, capability flags (tools/thinking/media), and model limits. It is consumed as structured configuration data, not as callable API code.
  • Where documented: Providers

src/doc_processing_toolkit/core.py

  • Coverage type: full API reference
  • Overview: Shared token-counting and cost-calculation functions used by provider call paths and orchestration flows. This module defines core accounting behavior for input/output/reasoning tokens, prompt-cache reads and writes, tiered pricing, and built-in tool costs. An agent's cost also adds the spend its tools report (utils.tool_spend_marker, Cost Tracking).
  • Where documented: Toolkit Core API

src/doc_processing_toolkit/multimedia.py

  • Coverage type: full API reference + concept page + recipe
  • Overview: One-shot text-to-speech, speech-to-text, and image generation/editing on OpenAI and Google, validated against the model registry's capability metadata before any API call and priced through the audio_output, per_minute, and multimodal_output schemas. Speech can also be streamed as it is generated (synthesize_speech_stream / asynthesize_speech_stream, 0.5.1). It also carries describe_media (one media Q&A, modality checked before the upload) and inspect_media (the same questions answered locally, with no network call).
  • Where documented: Toolkit Core API + Multimedia + Speech, Transcription and Media Questions

src/doc_processing_toolkit/image_session.py

  • Coverage type: full API reference + concept page + recipe
  • Overview: ImageSession / ImageTurn — multi-turn image generation and editing on OpenAI and Google, with a versioned envelope that survives the process that created it. Re-exported lazily as multimedia.ImageSession, so reference tooling should point at this module rather than the alias.
  • Where documented: Toolkit Core API + Image Sessions + Multi-turn Image Sessions

src/doc_processing_toolkit/audio_convert.py

  • Coverage type: full API reference + concept section
  • Overview: Optional speech-to-text container transcoding. It discovers a converter rather than requiring one, so the [audio] extra is never a dependency an installing application pays for.
  • Where documented: Toolkit Core API + Multimedia

src/doc_processing_toolkit/orchestration.py

  • Coverage type: brief overview
  • Overview: Implements higher-order generator/evaluator orchestration workflows (multi-step refinement loops) on top of provider structured-output functions, with loop-level token/cost aggregation and evaluation gating.
  • Where documented: this page (primary)
  • Status: pre-Agent pipeline code, deprecation-tracked. Not a public entrypoint — do not build new work on it; use the Agent API instead.

src/doc_processing_toolkit/pdf_processor.py

  • Coverage type: full API reference + recipes
  • Overview: End-to-end PDF preprocessing surface for directory traversal, PDF-to-image conversion, image preprocessing, and optional LLM Whisperer extraction workflows. It is a primary entrypoint for document ingestion pipelines.
  • Where documented: Toolkit Core API + PDF Processing Pipeline

src/doc_processing_toolkit/prompts.py

  • Coverage type: brief overview
  • Overview: Central prompt template catalog for image extraction, markdown refinement, engineering document reconstruction, and loop-based generation/evaluation flows. This module is primarily content/configuration rather than procedural API logic.
  • Where documented: this page (primary)

src/doc_processing_toolkit/realtime.py

  • Coverage type: full API reference + concept page
  • Overview: Provider-neutral realtime sessions over OpenAI Realtime and Gemini Live, wrapping each vendor SDK's own realtime client behind one connect/configure/send/events lifecycle with per-modality usage and cost tracked live. Besides speech-to-speech it carries visual input (stills and rate-gated video frames) and dedicated transcription sessions, routed by (provider, session_kind). The translation and GPT-Live families live in their own modules below and register themselves into the same routing table.
  • Where documented: Toolkit Core API + Realtime Voice

src/doc_processing_toolkit/realtime_translation.py

  • Coverage type: full API reference + concept section
  • Overview: Dedicated live-translation sessions — speech in, translated speech and both transcripts out, with no model turn and no turn-taking. GeminiLiveTranslateAdapter reuses the Gemini Live socket with a translation_config; OpenAITranslationAdapter is the module's one documented raw-WebSocket exception, because the openai SDK cannot address /v1/realtime/translations. Both register themselves into realtime's routing table.
  • Where documented: Toolkit Core API + Realtime Voice

src/doc_processing_toolkit/realtime_gpt_live.py

  • Coverage type: full API reference + concept section
  • Overview: OpenAI's GPT-Live family on its own /v1/live/* endpoints — a full-duplex voice session with no audio commit, no end-of-response event and no item ids, billed by the second rather than by tokens, whose tools run against a delegated backend. It goes through the openai SDK's own Live client and registers itself into realtime's routing table.
  • Where documented: Toolkit Core API + Realtime Voice

src/doc_processing_toolkit/schemas.py

  • Coverage type: full API reference
  • Overview: Defines shared Pydantic/dataclass schema objects used for structured output, evaluation workflows, and built-in tool configuration plumbing across providers.
  • Where documented: Toolkit Core API

src/doc_processing_toolkit/utils.py

  • Coverage type: full API reference
  • Overview: High-traffic utilities layer for provider client initialization, validation logic, media/citation handling, model-info helpers, and shared compatibility transforms used by both Agent and provider modules. It also holds the one retry layer (the should_retry_api_call predicate, the retry-after-aware wait strategy and the DPT_SDK_RETRIES client setting), the tool-result size cap and the reasoning-into-schema merge switch. Since 0.5.1 it also holds the tool-reported spend marker (tool_spend_marker, read_tool_reported_spend), the tool_call_result event fields (tool_result_event_fields, tool_result_data), the TruncatedResultStore behind read_truncated_result, and IterationEventUsage, which the four tool loops fill for iteration_complete.
  • Where documented: Toolkit Core API