Using Search to retrieve context in a prompt

Vellum Search can be used to include relevant context in a prompt at run-time that fits within LLM token window limits. Typically, the query that produces the search results comes from an end-user.

For example, here’s a prompt used to answer questions, pulling the source materials for an answer from a document index. The remainder of this article references variables from this prompt to explain the mechanics.

Answer questions based on the context provided below without
using any other knowledge. If the question can't be answered
using the provided context say "Sorry, I don't know."
Answer in the following format:
Question: ..
Answer: ..
Question: {query_str}

Step 1: How to add relevant context to context_str variable

In Vellum Playground, each variable has a 🔍 icon to include search results in Playground

Prompt Search Icon

Clicking this button opens up a modal to return Search results for a given query. Follow these steps in the modal:

  • Enter the Document Index that should be queried against
  • Type the search query (note: this should be the same as the query_str in the prompt variable)
  • Choose the number of chunks to be returned, the separators between each chunk and hit Run. By default, Vellum returns 3 chunks and uses new lines as separators.
  • Clicking Apply adds these results to the context_str variable

Prompt Search Dialog

Step 2: How to use the context to get results

Copy/paste the same user query into the query_str variable. Hit Run on the Playground and see the results against a variety of prompts.

Prompt Search Results