Prompt for page numbers
LLMs hallucinate citations when the contract is not enforced.
DocTruth
Open-source auditable LLM extraction for Java enterprise systems. Every extracted field carries source page + line, confidence, and bi-temporal provenance.
- Java
- 25+
- Maven
- 0.2.0-alpha
- License
- Apache 2.0
Enterprise AI dies at one question.
When a Java backend extracts a value from a document, DocTruth keeps the source location, match quality, confidence, and provenance attached.
String-match output back to PDFs
Brittle on tables, columns, scans, and formatting drift.
Glue callbacks onto LangChain4j
The app inherits framework coupling without evidence guarantees.
Run a Python service beside Java
Java enterprise teams reject the extra runtime topology.
Small primitives, auditable output.
A focused evidence layer that drops into Java backends already calling OpenAI-compatible endpoints, Anthropic, or Gemini.
Document parsing
PDF/DOCX/XLSX/CSV into structured sections with page, line, and offset preserved.
Evidence-attributed extraction
LLM calls are wrapped by a citation contract; missing evidence triggers validation and retry.
Smart context assembly
Priority truncation, sliding windows, and hierarchical context for over-window documents.
Fluent Java API
DocTruth.from(provider).extract(...).withProvenance().run(doc).
Audit export
PROV-O JSON-LD, confidence, retry count, model version, extracted_at, source_published_at.
Install in one coordinate.
Published on Maven Central. No framework runtime, no Python service, no extra deployment topology.
Maven
<dependency>
<groupId>ai.doctruth</groupId>
<artifactId>doctruth-java</artifactId>
<version>0.2.0-alpha</version>
</dependency> Gradle
implementation "ai.doctruth:doctruth-java:0.2.0-alpha"record Contract(String partyA, String partyB, BigDecimal totalValue) {}
var doc = PdfDocumentParser.parse(Path.of("contract.pdf"));
var result = DocTruth.from(new OpenAiProvider(System.getenv("OPENAI_API_KEY")))
.extract("Extract the contract terms", Contract.class)
.withProvenance()
.withSourcePublishedAt(Instant.parse("2026-01-01T00:00:00Z"))
.withBitemporal()
.withConfidence()
.run(doc);
Citation cite = result.citations().get("totalValue");
Confidence conf = result.confidence().get("totalValue");
result.toAuditJson(Path.of("audit/contract.jsonld"));A small explicit API for reconstructable extraction.
The value is not enough. The result also carries the exact quote, source location, match score, confidence rationale, model, version, and timestamps.
- Per-field citations: page, line, quote, match score
- Confidence scores with rationale, not only a number
- Bi-temporal provenance for source time vs extraction time
- Audit JSON that downstream systems can ingest
Non-silent by design.
A failed citation match is not allowed to vanish. It emits a warning and surfaces a low match score so the caller can decide how to handle risk.
Open-source core.
- Provider clients for OpenAI-compatible endpoints, Anthropic, and Gemini
- PDF/DOCX/XLSX/CSV parsing primitives
- Citation, Confidence, Provenance records
- Context strategies and audit JSON export
- Current SPI hooks for signing, audit events, and OCR
Built for Java teams.
- Parse source documents
- Extract typed Java records
- Validate schema and citations
- Export audit-ready JSON