DEV Community

Garfield Assay
Garfield Assay

Posted on

Your AI has a knowledge problem, not just a model problem

Enterprise AI often looks impressive in a controlled demo and becomes unreliable when it meets real company knowledge.

The model may be capable. The retrieval pipeline may work. Yet the source material still contains:

  • Several versions of the same policy
  • Documents whose owners have left
  • Fresh content mixed with outdated instructions
  • Duplicate files spread across multiple systems
  • Answers without traceable sources
  • Restricted knowledge that should not be exposed to every user

When these materials enter a retrieval system unchanged, the AI does not simply inherit the company’s knowledge. It inherits its contradictions.

RAG is necessary, but it is not the whole control system

Chunking, embeddings, hybrid search and reranking can improve retrieval. They do not determine which of two conflicting documents should be trusted.

That requires a governed knowledge supply chain.

Before an enterprise AI application or Agent uses a piece of knowledge, five controls should be clear:

1. Source connection

Where did the knowledge come from?

The connection should preserve source identity, permissions, timestamps and content type—not reduce everything to anonymous text.

2. Quality and authority

Is the material current, complete and owned?

Human verification, source authority, freshness and usage signals can help distinguish an approved policy from an abandoned draft.

3. Duplicate resolution

Are multiple documents repeating or contradicting the same knowledge?

Duplicate detection should go beyond filenames. It should identify semantically similar materials and apply explicit survivorship rules to determine the authoritative version.

4. Permissions and lineage

Who is allowed to use this knowledge, and how did it reach the answer?

Every output should remain traceable from its source through processing, classification and retrieval.

5. Human governance

What happens when the system is uncertain?

Low-confidence matches, conflicting sources and high-risk knowledge need a review queue with clear ownership. Human judgment should be part of the operating model, not an emergency fallback.

The practical question

Before asking, “Which model should we use?”, enterprises should ask:

Is our knowledge ready to be used by AI?

If the answer is unclear, improving the model alone will not create trustworthy outcomes.

We are building Assay to help enterprises diagnose and govern this layer before unreliable knowledge reaches their AI applications and Agents.

Take the 60-second AI Knowledge Readiness Check:

https://assay.20190628.xyz/en?leadPrompt=1&utm_source=devto&utm_medium=organic_community&utm_campaign=knowledge-readiness-pilot&utm_content=knowledge-supply-chain-post

Top comments (0)