AI Knowledge Assistant: internal document Q&A with source citations

The problem

Digitized documents sit in a system but are still hard to use. Keyword search requires exact terms and file codes; new staff depend on experienced colleagues; pulling an answer together from several documents takes hours.

The solution

The assistant retrieves relevant passages the user is permitted to see, then uses a language model to compose a concise answer with citations to the source document and page. When there is not enough evidence, it says so instead of guessing.

Workflow

1. InputDocuments, digitized records, PDFs, internal files
2. ProcessingExtraction, chunking, indexing, metadata and access rights
3. Q&APermission-aware retrieval, language model composes the answer
4. OutputAnswer with sources, original document one click away

Core features

• Natural-language Q&A in Vietnamese
• Document and page citations for every point
• Summaries and comparisons across documents
• Filters by unit, document type and date
• Respects existing document permissions
• User feedback loop to improve quality

Architecture

Four layers: data processing (OCR, extraction, chunking), search (full-text and vector indexes), language model (Anthropic's Claude via the Claude API) and integration (web interface, API, organizational sign-in). See the reference architecture.

Security

• Access checks before any content enters the model's context
• Audit log of questions, sources used and users
• Enterprise AI services that do not train on client data
• Deployment on infrastructure chosen by the organization

Current status

Pilot The assistant runs in real use on the Claude API, at pilot scale. TNC is taking on more organizations that want to pilot it on their own data.

Research

Under research and internal technical testing.

Prototype

Working prototype, demo by appointment.

Pilot

Being piloted with clients on real data.

Available

Ready for client deployment.