Engineering Note: Building a Navigable Semantic Data Store

Content type: Technical engineering note. This article describes document structure, retrieval and navigation rather than the full Semantic Labs insurance product proposition.

Why flat documents are difficult to use

Markdown and policy documents are readable to people, but their structure is often lost when they are treated as a flat stream of text. Headings, sections, dependencies and the scope of each passage matter when someone is trying to find and verify an answer.

A keyword search for “installation” or “exclusion” may return a relevant word without preserving the section, hierarchy or surrounding context that gives it meaning.

This note describes an engineering approach to preserving document structure so both people and AI systems can navigate information more deliberately.

From headings to a semantic tree

The first step is to treat headings as containers rather than visual markers. Each heading owns the content that follows it until another heading of equal or higher level appears.

That creates a semantic tree in which:

• A document has a root.
• Sections become nodes.
• Subsections become child nodes.
• Paragraphs, tables and lists remain associated with their relevant section.

The result is a navigable representation of the document rather than a flat text blob.

Why structure matters for policy intelligence

The same principle applies to insurance and regulated documents. A benefit may depend on a definition in one section, an exclusion in another, and an endorsement or schedule elsewhere.

Preserving structure can support:

• More focused retrieval.
• Better source context.
• Cross-section navigation.
• Change impact analysis.
• More reliable review of policy relationships.

A semantic data store is not, by itself, a policy decision engine. It is an information foundation that helps downstream reasoning systems locate and preserve the context needed for expert review.

Persistence and portability

A production document store needs to survive restarts, support multiple documents and provide predictable access patterns. A relational representation with parent-child links can provide a practical foundation while preserving a clear path to richer graph representations if the use case requires them.

The key design principle is not a particular database technology. It is retaining the identity, hierarchy, source location and relationships needed to reconstruct where information came from.

Summaries and navigation

Generative summaries can provide a useful overview of a section or document. They should be treated as navigation aids rather than authoritative answers.

A useful pattern is hierarchical summarisation:

1. Summarise the most specific sections.
2. Use those summaries to create a higher-level section view.
3. Preserve links back to the underlying content.
4. Retrieve the source section before relying on a conclusion.

This reduces context overload while maintaining a route back to the original text.

Agentic navigation

An AI agent can navigate a structured document by listing child sections, retrieving a node, moving back to a parent and selecting another branch. This is different from returning every paragraph that matches a keyword.

A navigation trace can show:

• Which sections were considered.
• Which source nodes were opened.
• Where the answer was found.
• When the agent backtracked.
• Whether the requested information was absent.

For regulated workflows, this trace is useful evidence of information retrieval. It is not a substitute for approved policy logic, evidence validation or authorised review.

The commercial connection

For life insurance, document structure is one foundation for reducing the PDF tax. It can help make policy documents easier to navigate and provide better source context for product comparison, claims review, underwriting support and AI-enabled customer conversations.

The larger Semantic Labs proposition adds structured policy logic, causal relationships, version control and human approval to that document foundation.

Conclusion

The engineering lesson is simple: before asking AI to reason over a document, preserve the structure that gives the document meaning.

A navigable semantic store helps people and AI find the right context. Governed policy intelligence then determines how that context can be compared, tested and used in a high-consequence workflow.