Overview
Approved answers with traceable sources
Every answer shows its source, unsupported cases are surfaced instead of guessed, and reviewed knowledge can be reused without pretending generated text is automatically trusted.
Project
Evidella
Building Evidella, a source-backed AI workspace for teams that need approved answers they can prove. It helps teams answer customer questions, security questionnaires, RFPs, internal Q&A, and repeated operational requests from approved company documents.
Problem
The problem
Customer questions, security questionnaires, RFPs, internal Q&A, and repeated operational requests create answer work that is repetitive but still high-risk.
Constraints
Constraints
Answers needed traceable evidence, unsupported requests could not pass as complete, and reviewed outputs had to stay reusable inside AI tools without opening uncontrolled write access.
Contribution
My work
I built the workflow across document ingestion, retrieval, structured answer generation, questionnaire review, missing-source handling, approval states, AI access patterns, and async processing.
System
Request flow
Approved documents
↓
Ingestion and normalization
↓
Retrieval over approved sources
↓
Structured answer generation
↓
Review and approval workflow
Technical decisions
- Keep retrieval constrained to approved company sources
- Separate unsupported answers from successful but imperfect answers
- Use structured outputs so reviews and downstream actions stay predictable
- Expose approved knowledge to AI tools through read-only access patterns
Result
What it improved
The result is a workspace where teams can reuse approved answers, review questionnaires faster, and connect reviewed knowledge to AI tools without treating model output as final truth.
FAQ
Common questions
How is retrieval constrained?
Retrieval runs over approved company documents so answers stay tied to sources the team is willing to stand behind.
What happens when the system cannot support an answer?
Unsupported questions are flagged, missing-source requests can be created, and the workflow stays explicit instead of inventing an answer.
How are questionnaires reviewed at scale?
Questionnaires, RFPs, and spreadsheets can be imported into review flows with reusable answers, review status, and approval steps.
Where do AI tools connect?
Approved answers can be exposed through read-only access patterns to tools like ChatGPT, Claude, Codex, and Cursor so teams can reuse reviewed knowledge without opening unrestricted write paths.