Retrieval Is Not Enough: AI for Organizations Needs Epistemic Infrastructure

Abstract

The problem: relevant, yet misleading

What is not yet established

• No longitudinal validation. The time-dependent node-priority dynamics are a specification, not a measured result.
• No dynamic-vs-static isolation. A static typed-graph baseline is the decisive comparison; dynamic weighting remains an open hypothesis.
• In-window regime. The full-context baseline sees the entire corpus and therefore wins aggregate quality in the reported setting.
• The graph must be built accurately. Typed graph retrieval only helps if nodes and edges are extracted with adequate precision and recall.

@article{bottino2026retrieval,
  title   = {Retrieval Is Not Enough: AI for Organizations Needs Epistemic Infrastructure},
  author  = {Bottino, Federico and Ferrero, Carlo and Dosio, Nicholas and Beneventano, Pierfrancesco},
  journal = {arXiv preprint arXiv:2604.11759},
  year    = {2026},
  url     = {https://arxiv.org/abs/2604.11759}
}

Frequently asked questions

What is OIDA?

OIDA organizes the information an AI uses to answer questions about an organization. Each record has a label, such as decision, evidence, plan, or open question. Connections show which records support, contradict, depend on, or replace others. Dates and priority scores help the system choose what still applies. Researchers describe this as a typed, attributed, signed, time-indexed graph.

What does “epistemic infrastructure” mean?

It is the part of an AI system that records the status of its information. For example, it can distinguish an approved decision from an untested idea, show which plan a later decision replaced, and keep track of unanswered questions. Those labels and connections stay in memory and can be used in later answers.

How is OIDA different from RAG?

Retrieval-augmented generation (RAG) gives an AI relevant passages to use when answering. Standard RAG often chooses them by similarity to the question. OIDA also considers each record's label, priority, and connections. It can lower an old plan's priority, retrieve both sides of a disagreement, or return an unanswered question as a record.

Does graph memory beat long-context prompting?

Not in the reported pilot. When the model could read the whole document collection, that approach scored higher for overall answer quality. Structured retrieval used about 42 times fewer input tokens on ClearPath and surfaced explicit knowledge gaps more consistently. Tokens are the units of text a model processes. The pilot does not establish that a graph gives better answers in general.

What are the OIDA evaluation corpora?

They are three synthetic document collections designed to resemble organizational work: ClearPath covers consulting and operations, FireGlass covers IoT product development, and Vertex Minds covers venture capital. They test whether a system can recover current decisions, disagreements, replaced claims, and unanswered questions. The authors released them publicly so others can repeat the tests.

When is epistemic graph memory worth using?

Consider it when the relevant documents cannot fit affordably into the model's input and an answer depends on changing decisions, conflicting sources, open questions, differing authority, or connections across documents. If reading the full collection is affordable, use that. Ordinary RAG may be enough for a fact found in one source.