WOFO AI
Smart RAG .Secure answers from governed documents.
Created on 30th December 2025
•
WOFO AI
Smart RAG .Secure answers from governed documents.
The problem WOFO AI solves
Most AI assistants are designed to always answer, even when the information is incomplete, outdated, or restricted.
This becomes dangerous in enterprise environments where permissions, compliance, and accuracy matter.
This project solves the problem of unsafe and uncontrolled knowledge access by introducing governance into Retrieval-Augmented Generation (RAG).
What people can use it for
Enterprise knowledge access
Safely query internal documents such as HR policies, technical documentation, or internal reports.
Role-restricted information retrieval
Ensure employees only see information they are authorized to access based on their role.
Compliance-sensitive environments
Prevent leakage of confidential or regulated information by enforcing document-level and chunk-level permissions.
Auditable AI decisions
Track who queried what, why an answer was given, or why a query was rejected.
How it makes existing workflows safer
Prevents hallucinated or unauthorized answers
Refuses to answer when information is missing or restricted
Makes AI outputs traceable, explainable, and confidence-scored
Separates knowledge governance from language model generation
Instead of behaving like a chatbot, the system acts as a controlled knowledge layer — prioritizing safety and trust over convenience.
image related to clusters :-
Challenges we ran into
- Preventing the system from “over-answering”
One major challenge was ensuring the system did not generate answers by default.
Early versions returned responses even when retrieved context was weak or incomplete.
How I solved it:
Introduced strict pre-LLM filtering
Added rejection logic when no permitted or relevant chunks exist
Logged all rejections for auditability instead of silently failing
- Managing permissions across documents and chunks
Applying permissions only at the document level was too coarse and unsafe.
Some documents contained sections that should be restricted even if the rest was allowed.
How I solved it:
Introduced chunk-level metadata with allowed roles
Ensured permission checks happen before passing context to the LLM
Kept governance logic in the database layer to avoid backend coupling
- Designing auditability without breaking existing logic
Adding analytics, query logs, and confidence tracking without refactoring backend code was challenging.
How I solved it:
Designed append-only MongoDB collections for logging
Treated the database as a control plane, not just storage
Added observability without changing execution paths
- Avoiding tight coupling between systems
Another challenge was preventing hard dependencies between MongoDB, Qdrant, and the LLM.
How I solved it:
Clearly separated responsibilities:
MongoDB → governance & audit
Qdrant → semantic retrieval
LLM → response generation
Ensured each layer can evolve independently
Tracks Applied (2)
Best Innovation
AWS
AWS