Structured lookup and text fallback

The project exposes two complementary evidence paths.

Structured lookup

Structured resolvers query imported registries, tables, and graph relationships.

Use them for:

• tags, keywords, VR, and VM;
• UIDs and retirement state;
• IODs, modules, macros, and attribute uses;
• SOP Classes and linked IODs;
• value terms;
• VR and Transfer Syntax details;
• DICOMweb transactions and media types;
• SR templates, context groups, and code meanings.

Benefits include stable fields, deterministic validation, candidate handling, and high or medium parse confidence.

Cited text retrieval

Text tools search or retrieve persisted DocBook prose.

Use them for:

• a rule without a dedicated semantic parser;
• selected PS3.7 or PS3.8 messaging and networking topics;
• explanatory context around a structured result;
• inspecting an exact section or anchor.

Text results are bounded, reference preserving, and deliberately classified as explanatory with low parse confidence.

Preferred evidence ladder

1. exact structured resolver
2. graph or contextual resolver
3. specialized bounded explanation
4. part-filtered full-text search
5. bounded section retrieval
6. explicit statement that the fact is unsupported or ambiguous

Model memory is not a fallback for normative DICOM facts.

Why not search first?

A text match can:

• lack the surrounding table;
• refer to a different context;
• be explanatory rather than normative;
• use a phrase that appears in multiple parts;
• omit a condition linked through another structure.

A specialized resolver can anchor the investigation and identify the correct references before prose is retrieved.

Why text still matters

Not every standard rule is naturally reducible to a table row. Prose can explain:

• conditional behavior;
• encoding rationale;
• protocol sequencing;
• implementation constraints;
• exceptions and notes.

The project preserves text retrieval as evidence without pretending it is fully structured.

Synthesis rule

A good answer separates its components:

Parsed fact: ...
Structured context result: ...
Supporting excerpt: ...
Unresolved condition or warning: ...
Official references: ...
Edition and source-manifest trace: ...

This structure lets a reader verify the source and see exactly where machine interpretation stops.