THE PROBLEM

A general-purpose chatbot cannot be trusted with policy

Agencies hold the answers in writing — thousands of pages of regulation, guidance, memoranda and procedure. Finding the right passage takes an experienced person, and that person is the bottleneck.

A general model will answer confidently from whatever it absorbed during training, including superseded guidance and guidance that was never yours. In a regulated process, an uncited answer is not an answer. It is an exposure.

The constraint we build in
Answers come from your corpus only

Retrieval is restricted to the documents you load. If the corpus does not cover a question, Doktrin says so rather than filling the gap.

Every claim carries a citation

Each response links to the document, section and passage behind it, so verification takes seconds.

Permissions are honored at retrieval

A user only ever gets answers drawn from documents they are already cleared to read.

CAPABILITIES

What Doktrin does

Grounded retrieval

Semantic search across your document corpus, returning the passages that actually answer the question rather than the ones that share keywords.

Source citation

Every response names its sources and links to the exact passage, with the document version and effective date attached.

Version awareness

Distinguishes current guidance from superseded guidance so staff are not quoting a rescinded memorandum.

Access-aware answers

Retrieval respects existing document permissions, so the agent never becomes a path around your access controls.

Conversation history

Follow-up questions keep context, so staff can work a problem rather than re-ask it five ways.

Usage analytics

Shows which questions staff ask most and where the corpus has gaps, so guidance owners know what to write next.

DEPLOYMENT

From kickoff to production

01
Define the corpus

We identify the authoritative sources and who owns each one.

02
Deploy in your boundary

Installed on premises or in your authorized cloud region, wired to your identity provider.

03
Evaluate on real questions

Scored against a question set your subject matter experts write and grade.

04
Run and sustain

Prudent keeps the corpus current and operates it under your service level.

SECURITY

Nothing leaves the boundary

  • Models and retrieval run inside your authorization boundary; no query or document is sent to an external service.
  • Agency content is never used to train or tune models.
  • Document-level permissions enforced at retrieval, federated with your identity provider.
  • Full audit log of questions asked, sources returned and users served.
  • Air-gapped deployment supported for classified and restricted environments.
SPECIFICATIONS

Technical summary

Deployment
On premises, agency data center, authorized government cloud, air-gapped
Source types
PDF, Office documents, HTML, wikis, SharePoint, ticketing and case systems
Interfaces
Web chat, REST API, embeddable widget, voice
Identity
SAML and OIDC federation with agency identity providers; PIV/CAC supported
Answer quality
Measured on a customer-authored evaluation set during pilot
Languages
Major spoken languages supported (English, Spanish, French, Hindi, Chinese, Arabic, Russian & many others)
Licensing
Annual subscription, multi-year discounts available

Test it on the questions you already get

Give us your guidance library and twenty questions your team fields every week. We will show you the answers and the citations behind them.