Best Practices and Costs

Build a focused knowledge base

Prefer fewer, higher-signal documents over large dumps of loosely related files.
Keep sources up to date: outdated SOPs will produce outdated answers.
Use clear structure: headings, consistent section names, and short paragraphs improve retrieval.

Use text-friendly formats when possible

MD/TXT are typically easiest to ingest and retrieve from.
If you use PDF, ensure it contains selectable text (scanned PDFs often ingest poorly unless OCR’d).

Write good prompts & instructions

Your assistant instructions should describe:

Role (who it is): “You are a lab operations assistant…”
Scope (what it can answer): “Only answer using the uploaded SOPs…”
Output format (how to respond): “Return concise steps and list assumptions.”

Knowledge Builder responses are designed to be plain text. If you need a specific structure (for example, numbered steps), explicitly request it.

Cost considerations

Knowledge Builder can incur costs in two places (billed by your AI provider):

Ingestion: when you upload sources, Foundry Connect extracts content and generates embeddings.
Q&A: each question uses tokens for retrieval + the final model response.

PDF/HTML with images can cost more

PDF and HTML sources may include charts or images. During ingestion, Foundry Connect may analyze visuals to improve retrieval quality, which can increase token usage and ingestion time.

Manage costs with model selection

Use a lighter / faster model for iterative Q&A when you don’t need deep reasoning.
Use a stronger model when correctness and nuanced interpretation matter.