raja-jaloka/veriaudit-

Purpose and Entry Points

VeriAudit is a local full-stack document-intelligence system that audits text against compliance rules, extracts grounded facts, synthesizes domain rules, and rejects prompt injection. The backend FastAPI application starts at backend/main.py; the Vite/React workbench starts at frontend/src/main.tsx and mounts frontend/src/App.tsx. Batch verification enters through run_evals.py.

API Orchestration Boundary

backend/main.py is the server-facing boundary for audit, ad-hoc rule prompting, rule synthesis, entity extraction, and file upload APIs. It accepts document text or uploaded files, invokes deterministic engine stages, and returns typed response payloads defined in backend/models/schemas.py. The frontend consumes this API rather than embedding audit logic locally.

Document Ingestion and Canonical Text

Uploaded source files are converted into a canonical document-text buffer before analysis. backend/ingestion/file_parser.py handles supported document formats; backend/ingestion/cleaner.py normalizes content; backend/ingestion/chunker.py prepares analysis-sized text segments. Character-position fidelity at this boundary is architecturally critical because downstream citations refer back to the original text buffer.

Grounded Audit Pipeline

The audit path evaluates selected or supplied rules against document text. Rule packs originate in backend/rules/presets.py; backend/engine/audit_agent.py coordinates verification. backend/engine/grounding_validator.py validates affirmative and counter-evidence as exact document substrings and character offsets, establishing the no-unsupported-claim boundary. backend/engine/hallucination_guard.py screens adversarial or injection-style instructions before they can alter rule interpretation.

Rules and Ad-Hoc Queries

Rules can be loaded from presets, dynamically generated, or derived from a user question. backend/engine/rule_synthesizer.py classifies document vocabulary/domain and creates tailored rule candidates. The same API orchestration in backend/main.py exposes natural-language prompt rules, translating a question into a form evaluated by the grounded audit path rather than returning an unconstrained model-style answer.

Fact Extraction and Confidence

backend/engine/extractor.py identifies structured entities and contextual occurrences across domains, including repeated mentions that may be consolidated into canonical records. backend/engine/confidence_engine.py computes evidence-based confidence from grounding, salience, multiplicity, authority, and penalties. These outputs augment audit reports but remain subordinate to substring validation.

UI Workbench

The React UI is an operator surface for supplying documents, selecting/managing rules, running audits and benchmarks, and inspecting/exporting evidence. frontend/src/components/ImportDocumentModal.tsx manages intake; frontend/src/components/AdHocPrompter.tsx drives query-based audits; frontend/src/components/AuditReportView.tsx presents verdicts and citations. frontend/src/components/SplitPaneDocumentViewer.tsx is the key evidence-review boundary between source text and results.

Evaluation and Runtime

Python 3.10+ runs FastAPI/Uvicorn; Node 18+ runs the Vite frontend. The evaluation harness in backend/evaluator/runner.py and backend/evaluator/benchmark_suite.py executes fixed benchmark scenarios, with run_evals.py as its root launcher. The repository indicates no database, queue, cloud runtime, or external AI service dependency; analysis is organized as in-process Python engine modules.