Package ai.doctruth

package ai.doctruth

Public API of DocTruth: auditable LLM extraction for Java.

Every type in this package (and only this package) is part of the stable public API. Subpackages under ai.doctruth.internal are explicitly NOT public API and may change without a major version bump.

See CONTRIBUTING.md in the repository root for the engineering contract.

Since:

0.1.0

Related Packages

Package	Description
ai.doctruth.cli
ai.doctruth.internal	Internal implementation details.
ai.doctruth.spi	Service Provider Interface (SPI) extension points for DocTruth.

Class	Description
AnthropicProvider	Anthropic Messages API provider.
BlockKind	Visual / structural classification of a TextSection as detected by Layer 1 parsing — a geometric / typographic judgement, NOT a semantic one.
Citation	The verifiable evidence anchor for a single extracted field.
Confidence	A confidence score for a single extracted field, plus a free-form rationale.
ContextStrategy	How a ParsedDocument is rendered into the bounded context window of an LLM call.
CsvDocumentParser	Layer 1 entry point: read a CSV file from disk into a ParsedDocument containing exactly one TableSection that mirrors the CSV row-major.
DeepSeekProvider	DeepSeek Chat Completions provider.
DocTruth	Public entry point for the library.
DocumentMetadata	Metadata for a ParsedDocument: the source filename, total page count, and (optionally) the timestamp at which the source document was authored / published.
DocxDocumentParser	Layer 1 entry point: read a DOCX file from disk into a ParsedDocument with one TextSection per non-blank paragraph and one TableSection per table.
ExtractionBuilder<T>	Immutable fluent builder for one extraction call.
ExtractionException	Thrown from the public extraction API when an extraction run fails after exhausting retries or when an invariant is violated mid-flight.
ExtractionResult<T>	The output of DocTruth.extract(...).run(): the extracted value plus per-field citations, per-field confidence scores, and run-level provenance.
FigureSection	A figure (image, chart, diagram) recovered from the source document, represented by its caption text plus a SourceLocation.
GeminiProvider	Google Gemini generateContent provider.
Hierarchical	Multi-level summarisation: condense the document at increasing granularities, hand the LLM the level that fits the budget.
JsonExtractionBuilder	Immutable fluent builder for JSON Schema-driven extraction.
JsonSchema	Caller-supplied JSON Schema for schema-bound extraction.
LlmProvider	The Layer 2 backend abstraction: an LLM API client.
OpenAiProvider	OpenAI Chat-Completions API provider.
OverBudgetPolicy	What a PriorityTruncate strategy does when the priority sections alone exceed the configured maxChars budget.
ParsedDocument	The output of Layer 1 (document parsing): a stable identifier, an ordered list of sections, and the document metadata.
ParsedSection	A single section of a parsed source document.
ParseException	Thrown by Layer 1 document parsers (PDF / DOCX) when a source file cannot be parsed or when a structural invariant is violated.
PdfDocumentParser	Layer 1 entry point: read a PDF file from disk into a ParsedDocument with source locations preserved per detected layout block.
PriorityTruncate	Smart-context strategy for keeping priority sections while trimming everything else to fit.
Provenance	Bi-temporal provenance for an ExtractionResult: the model that produced it, when the extraction ran, and (optionally) when the source document was authored, the region the extraction was processed in, and the retention horizon of the audit record.
ProvenanceDetails	Supplemental provenance metadata kept behind Provenance so the public provenance record stays small while preserving retry, data-residency, and retention semantics.
ProviderException	Thrown by Layer 2 LLM providers (Anthropic, OpenAI, Gemini, DeepSeek) when an upstream call fails.
ProviderOptions	Per-call knobs passed to an LlmProvider on every request.
ProviderRequest	What the library hands an LlmProvider on every call: the system prompt, the user prompt (rendered from a ParsedDocument by the configured ContextStrategy), the JSON Schema for the target type, and the per-call options.
ProviderResponse	What an LlmProvider returns on a successful call: the raw JSON the LLM produced plus the per-call ProviderUsage.
ProviderUsage	Token-usage and model-version data returned by an LLM provider on every successful call.
SlidingWindow	Fixed-size character windows with optional overlap.
SourceLocation	A 1-indexed page + line span into a parsed source document, plus a 0-indexed character offset into the source page text.
TableSection	A flat string-cell table recovered from the source document, anchored to a SourceLocation.
TextSection	A run of plain text recovered from the source document, anchored to a SourceLocation and tagged with a BlockKind that classifies the geometric / typographic shape of the block (HEADING / BODY / LIST / OTHER).
XlsxDocumentParser	Layer 1 entry point: read an XLSX file from disk into a ParsedDocument with one TableSection per non-empty sheet.