Enterprise document intelligence
From raw documents to decisions with a source.
A bilingual document intelligence platform that connects extraction, reviewed facts, policy evidence and bounded AI investigation. Every authoritative outcome has an explicit path back to its source.
View sourceThe problem
Business documents contain useful facts, but extracting a plausible value is only the beginning. A policy decision also depends on the correct source, current permissions, reviewed data and the meaning of the applicable rule. English and Arabic inputs add parsing, normalization and OCR challenges.
NEXUS separates proposing information from granting it authority. Models extract candidates and investigate with bounded tools. Shared eligibility checks, explicit human decisions and deterministic Python govern the official result. Source quotations and audit records let a reviewer reconstruct that decision.
An individual engineering project spanning a Next.js interface, FastAPI services, persistent workflows, retrieval, authorization and evaluation. Designed and verified in a local development environment.
Explore the system
Select a component to inspect its contract.
The workflow summarizes the end-to-end lifecycle. Policy ingestion and approval can happen independently. Restricted specialists investigate, while canonical data and verified rules determine the final evaluation.
Engineering challenges
Changing a source, reviewed field or permission can invalidate an earlier result. Workflow fingerprints, current source validation and resource checks on resume prevent historical approval from silently becoming current authority.
Database-backed jobs use expiring claims, bounded attempts and completion fencing. Business writes and job completion commit together, so a worker holding an expired claim cannot publish a stale result.
Unicode and RTL rendering are distinct from reliable Arabic OCR and semantic interpretation. Search normalization preserves original quotations, while OCR and extraction quality have separate development evidence and outstanding release gates.
Implementation & trade-offs
SQL scopes the candidate passages to the current organization, policy and enabled index before BM25 or vector ranking. Citation reads revalidate source hashes and offsets.
Trade-offThis maintains a clear access boundary. Exact vector scans and the 5,000-candidate bound limit scale. Scoped ANN retrieval is a future option when measurements justify it.
Keyword retrieval is the default. A pinned local multilingual ONNX encoder adds 384-dimensional vectors, cosine ranking and reciprocal rank fusion with BM25.
Trade-offOptional model assets keep the basic path lightweight. Semantic retrieval adds memory and model-version management, while better recall still does not establish answer correctness.
LangGraph coordinates separate Document and Policy specialists with disjoint tools. SQLite checkpoints persist review pauses, and canonical facts are reloaded before evaluation.
Trade-offA single-host checkpoint store keeps the system understandable. Distributed orchestration, retention and production recovery still require further work.
Search passages support investigation. The deterministic evaluator independently loads current, source-verified approved rules and reviewed document fields.
Trade-offA narrow supported monetary grammar favors inspectable behavior. Unsupported wording stays review-gated, and field review does not itself approve an expense.
Results & evaluation
Historical repository results.
Methods and limitations travel with the numbers.
32 source policies, 40 queries split equally between English and Arabic, and 8 unanswerable queries. Each answerable label includes both translated sources. Recall@5 measures recovery of both.
| Method | Recall@5 | MRR@5 | Abstention |
|---|---|---|---|
| Keyword | 50.0% | 0.8333 | 87.5% |
| Semantic | 79.7% | 0.8750 | 87.5% |
| Hybrid | 87.5% | 0.9635 | 75.0% |
Hybrid improves recall here, while unanswerable abstention drops to 75%. Arabic abstention is 50% on four unanswerable queries. These are small development samples.
Limits & lessons
Reliability comes from the boundaries around the model. A shared review predicate, current source checks and fenced state transitions are as important to an explainable result as the prompt or retrieval strategy.
Inspect the work
Selected implementation and documentation reviewed for this case study.