# Internal knowledge assistants

Build answers around retrieved passages, technical terminology and traceable document sources.

## Start with traceable documents

Prepare the source documents, their versions and the access rules for each reader. Technical handbooks may contain formulas, tables, abbreviations and references to other sections. Preserve page and section identifiers during parsing so answers can cite the material they use.

The result should be an answer supported by retrieved passages, or an explicit statement that the available material does not answer the question.

## Build the retrieval workflow

1. Parse the documents and check a sample of tables, formulas and headings against the originals.
2. Choose a retrieval method suitable for the material. Search, embeddings and a document store are separate application components.
3. Apply the reader's access rules before retrieving passages.
4. Send the question and relevant passages to the text model. Ask it to distinguish supported facts from missing information.
5. Return the answer with links or identifiers that lead back to the actual source passages.

## Test difficult questions

Include acronyms, references across chapters, conflicting document versions and questions with no answer in the corpus. Check that cited passages support the claim rather than merely contain similar words. A larger context window cannot repair badly parsed documents or missing retrieval results.

Use a fixed evaluation set when adding documents or changing retrieval settings. Track whether errors originate in parsing, retrieval or the final answer before changing the model.

## Match deployment requirements

Confirm where documents may be processed and whether the organization requires a particular model, cloud or offline environment. A hosted API may not fit every requirement. SeedRouter supplies model calls; it does not automatically provide a document store, embedding pipeline, access-control system or offline deployment.

Explore [text models](https://seedrouter.ai/text-generation), the [context-window guide](https://seedrouter.ai/blog/context-window-vs-max-output-tokens) and [current prices](https://seedrouter.ai/pricing). Account for document preparation and retrieval costs separately from the final model request.

## Related resources

- https://seedrouter.ai/models/gpt-6-sol
- https://seedrouter.ai/models/claude-opus-5-5
- https://seedrouter.ai/integrations/hermes-agent
- https://seedrouter.ai/industries/energy
- https://seedrouter.ai/blog/context-window-vs-max-output-tokens