A Vector Database Is Not Process Safety Memory

Most industrial organizations do not lose their process-safety documents. They lose the understanding that gave those documents meaning.

The HAZOP worksheet remains. The LOPA report remains. The recommendation is marked complete. The management-of-change record is closed. But years later, the team can no longer reconstruct why a scenario was considered credible, why a safeguard was credited, which assumptions supported the decision, what alternatives were rejected or what evidence was available at the time.

This is often described as a document-management problem. More recently, it has been presented as an opportunity for retrieval-augmented generation: place the documents in a vector database, add a conversational interface and allow engineers to ask questions.

That can make information easier to find. It does not create process safety memory.

Recent research across document intelligence, engineering diagrams, safety reasoning and AI governance points toward a more demanding architecture. The emerging lesson is that trustworthy engineering intelligence requires original evidence, structured relationships, explicit reasoning, change history and accountable human judgment.

For process safety, the unit of value is not merely the document or the answer. It is the connected engineering decision.

Search is useful, but search is not memory

A vector database represents passages according to semantic similarity. This is valuable when terminology varies or when the user does not know which document contains an answer. It can retrieve references to a safeguard even when different reports use different wording.

But similarity does not preserve the engineering structure around that safeguard.

Which hazardous scenario the safeguard applies to
Whether it was credited in a LOPA
Which equipment, tags and operating modes define its applicability
What performance assumptions were used and whether they remain valid
Which source and revision support the assertion
Whether a later MOC, incident or test result changed the basis
Who reviewed and accepted the decision

A retrieved paragraph may be relevant while still being incomplete, superseded or inapplicable to the decision at hand. Process safety is a governed decision environment in which evidence, context and applicability determine whether an answer can be relied upon.

What the recent research establishes

The papers do not provide a finished blueprint for Process Safety Intelligence. They do, however, reinforce several important principles.

Preserve the source and the relationships

The Sanofi-affiliated paper From Document Silos to Process Intelligence, published September 10, 2026, describes a dual-layer architecture applied to 38 reports from a process-development program.

One layer preserves the documents, sections, chunks, tables, figures and reading order. A second, ontology-driven layer connects domain entities across documents and versions. The relationships in the domain graph resolve back to their supporting passages.

The reported evaluation used 505 corpus-specific questions. Multiple-choice accuracy reached 95%, while stricter free-form evaluation reached 85%. Comparative and corpus-wide questions remained difficult. The authors report 98.9% direct or one-hop provenance coverage.

What the source shows: Combining lossless document representation with structured domain knowledge can improve retrieval while retaining strong links to source evidence.

ProsIQ inference: HAZOP and LOPA knowledge need two connected layers. The first preserves the record exactly as it existed. The second represents hazards, causes, consequences, safeguards, assumptions, decisions and actions as governed, versioned objects. Neither layer is sufficient by itself.

Recognizing a hazard is not the same as defending a conclusion

SafeSceneReason, published August 10, 2026, uses scene and evidence graphs to connect workplace observations with accident-investigation knowledge. Its report-centered approach establishes information boundaries, constructs multi-step reasoning paths and applies iterative verification.

The benchmark contains more than 110,000 verified scene-centered questions and more than 13,000 refined report-centered questions. Evaluated vision-language models continued to struggle with technical comparison, evidence synthesis and reasoning across multiple pieces of evidence.

What the source shows: Hazard recognition and defensible safety reasoning are different capabilities. Performance weakens when a conclusion depends on comparison, synthesis and multiple evidence sources.

ProsIQ inference: An AI-generated answer should not be treated as a reasoning record. A defensible record must show the question, information boundary, evidence considered, conflicting or missing evidence, reasoning path, verification status and human disposition.

Engineering diagrams should become structured evidence

ChatP&ID, published March 23, 2026, converts structured DEXPI P&IDs into knowledge graphs before enabling language-model interaction. The authors report an 18% accuracy improvement over direct diagram-image processing and an 85% reduction in token consumption. They identify AI-assisted HAZOP as a future application.

What the source shows: A structured representation of engineering topology can outperform treating the diagram as an image alone.

ProsIQ inference: A P&ID should be an active evidence source connected to HAZOP nodes, equipment, lines, instruments and safeguards. The immediate value is not autonomous HAZOP generation. It is the ability to identify discrepancies, changed topology and unsupported study statements for qualified engineers to review.

Probabilistic extraction needs deterministic provenance

Semi-Automated Knowledge Engineering and Process Mapping, published March 27, 2026, combines expert-defined symbolic structures with probabilistic language-model extraction. Extracted information is then anchored deterministically to its original source. The work also reports that whole-document processing recovered non-linear procedural dependencies better than isolated segment processing.

What the source shows: Domain structures can guide probabilistic extraction, while source anchoring can remain deterministic. Important dependencies may be lost when documents are processed only as isolated chunks.

ProsIQ inference: AI may propose that a statement represents a cause, safeguard or assumption. The system should preserve the exact source language and location, distinguish proposals from validated records and require controlled review before the information becomes governed knowledge.

“Human in the loop” is not a complete control

Generative AI in Systems Engineering: A Framework for Risk Assessment, published February 4, 2026 and accepted at IEEE SysCon 2026, classifies engineering AI applications using two dimensions: autonomy and the consequence of an incorrect output. The resulting risk class determines the validation, oversight and countermeasures required.

What the source shows: AI assurance should reflect both what a system is allowed to do and the consequence of it being wrong.

ProsIQ inference: Extracting an equipment tag, identifying a potentially similar scenario, recommending an IPL and accepting residual risk cannot share one generic review control. Each requires a different assurance level, evidence requirement, reviewer role and approval path.

What process safety memory actually requires

1. Lossless evidence

The original HAZOP worksheets, LOPA records, P&IDs, procedures, MOCs, incident reports, test records and meeting evidence must remain intact, identifiable and version-controlled. Structured knowledge should add meaning without replacing the record from which that meaning was derived.

2. Connected engineering objects

Hazards, scenarios, equipment, safeguards, assumptions, recommendations, actions and decisions must exist as connected objects rather than isolated text fragments. Their relationships must be explicit enough to examine, govern and challenge.

3. Decision context

A conclusion without its basis is an incomplete record. Process safety memory must retain why a decision was made, which evidence supported it, which alternatives were considered, which assumptions were material and who accepted the outcome.

4. Change intelligence

Revalidation must do more than retrieve the previous study. It must establish what changed in the equipment, process, procedures, staffing, standards, operating experience and evidence—and then identify which earlier conclusions depend on those changes.

5. Inspectable reasoning

AI assistance must expose its evidence boundary and reasoning path. Engineers need to see whether an answer is supported, contradicted, incomplete or dependent on an unverified assumption. A fluent explanation is not a substitute for an evidence chain.

6. Risk-tiered governance

The controls applied to AI should rise with autonomy and consequence. Low-impact extraction can use confirm-and-save review. Recommendations involving safeguard credit require stronger evidence and qualified approval. Final risk acceptance remains an explicitly human decision.

From periodic studies to continuous intelligence

Traditional process-safety systems are organized around activities and documents. A HAZOP is performed. A report is issued. Actions are tracked. Several years later, a team attempts to reconstruct what has happened since.

Process Safety Intelligence starts from a different premise: engineering reasoning is a lifecycle asset. HAZOP, LOPA, SIL, MOC, incident investigation, operating experience and design review should contribute to a persistent knowledge layer. Each new event should be able to strengthen, challenge or invalidate what the organization previously believed.

This does not replace established process-safety methods. It makes the knowledge produced by those methods persistent, connected and reusable. Nor does it transfer accountability to AI. AI can help retrieve, compare, structure and challenge information. Qualified people must continue to validate evidence, exercise judgment and approve consequential decisions.

The ProsIQ response

ProsIQ is being developed as one implementation of this broader Process Safety Intelligence model. Its direction is guided by a simple principle: Every Decision Connected.

That means connecting a process-safety decision to its reasoning, assumptions, evidence, ownership and lifecycle history. It means preserving the original record while building a governed knowledge layer across HAZOP, LOPA, MOC, incidents, safeguards, actions and operating experience. It means designing AI explanations as inspectable evidence chains rather than conversational assertions.

The objective is not to generate more HAZOP text. It is to help organizations preserve what their engineers knew, understand what has changed and make the next decision with the full context of the decisions that came before it.

A vector database can help find the documents. Process safety memory must preserve the thinking.