# Energy

Work with long technical documents, formulas and specialist terminology while keeping sources visible.

## Preserve technical context

Energy documentation can span long manuals with formulas, abbreviations, component names and references between sections. A useful assistant needs the right source material and a reliable way to return readers to it.

Prepare documents with revision identifiers, page numbers and access rules. Inspect parsed tables and formulas against the originals before evaluating model answers.

## Build sourced technical answers

The [knowledge assistant workflow](https://seedrouter.ai/solutions/knowledge-assistant) covers document preparation, retrieval and answer review. Use a glossary for domain terms, and retrieve the surrounding context when a passage refers to a different chapter or document.

A response should identify the supporting section and acknowledge when the material does not resolve the question. Similar wording in a passage is insufficient if the passage concerns a different component or operating condition.

## Evaluate specialist questions

Build a reference set with acronyms, formula definitions, cross-references and conflicting revisions. Ask a qualified reviewer to check whether each answer is supported. Diagnose parsing, retrieval and answer-generation errors separately.

Long context can help only when the relevant and correctly parsed material is actually provided. The [context-window guide](https://seedrouter.ai/blog/context-window-vs-max-output-tokens) explains the distinction between input capacity and output limits.

## Set the system boundary

Your organization controls document access, retrieval and professional review. SeedRouter supplies text model calls; it does not validate engineering decisions or operate equipment. Confirm model and deployment requirements before selecting a hosted API.

Compare costs using representative questions and document excerpts with [current rates](https://seedrouter.ai/pricing). Reuse the same evaluation set when updating manuals or changing a model so improvements in one question type do not hide regressions in another.