Skip to main content

​
“You Might Also Like”

A good answer is often only the start. Someone who asks how to talk with their teenager about faith may want a book on the subject, a sermon series or a podcast episode to listen to on the drive home. Recommendations suggest whole items, such as books, articles, videos and podcast episodes, that match what a person is asking about. They’re how you build “recommended reading” panels, related-resource lists and “you might also like” sections. On Gloo, recommendations come from content in the Data Engine, the part of Gloo that stores and indexes your uploaded content (AI 103 covered how that works). You’ll need content there, and Data Engine access, to use your own publisher’s items.

​
Core Concepts

​
Items and Snippets

When content goes into the Data Engine, each item is split into snippets: meaningful sections that can be searched by meaning. A book might become hundreds of snippets; a short article, a few. The Recommendations APIs search the snippets for your query, then group the matches by the item they came from. Items are ranked by:
  1. How many relevant snippets they have. An item with ten matching sections ranks above one with two.
  2. Cumulative certainty, as a tiebreaker. Each snippet has a certainty score saying how well it matches the query; cumulative certainty is the sum for the item.

​
Three Endpoints

The two item endpoints only return items your organization has access to. The affiliate endpoint is for pointing people to resources beyond your own library, and it never returns snippet text.

​
Recommendations, Search or a Grounded Answer?

All three search your content by meaning, but return different things:
  • Search returns the matching passages themselves.
  • Grounded Responses returns an answer written from those passages.
  • Recommendations return the items worth opening next.
Many apps use two together: a grounded answer, then a short list of recommended resources under it.

​
Get Started

You’ll need your tenant, which is your publisher’s name. Find it on the Organizations page in Gloo AI Studio under View Publishers. The collection is always GlooProd.
Python
Each item in the response looks like this:
To show a preview of each item, call the /verbose endpoint instead. It adds the snippet text to each entry in uuids, and lets you filter and sort, for example newest first with "sort_by": "publication_date". To recommend from the wider publisher network, call the affiliate endpoint. It doesn’t take a collection or tenant:
cURL
Affiliate items include display fields such as item_title, author, item_url and item_image, ready for a resource card.

​
Good to Know

  • No results? Lower the threshold. certainty_threshold filters out weaker matches. If you get nothing back, try a lower value, and test the same query in the Search Playground to confirm your content is indexed.
  • Filter by format. media_types limits results to audio, video, book or article.
  • Ask for fewer, better items. item_count caps the list, and min_snippet_count_per_item drops items with only a passing mention.
  • New network content takes time. Once a publisher allows their content to be referenced, it can take up to 24 hours to appear in results.
  • Stay in your own tenant. The item endpoints return 403 for a tenant your organization doesn’t own.

​
Key Takeaway

Recommendations point people to whole resources, ranked by how many relevant snippets each one has. Use /recommendations/base for clean lists from your own content, /recommendations/verbose for previews, and the affiliate endpoint for resources from across Gloo’s publisher network.
Try it on Gloo: the Building Custom Recommendations tutorial builds a recommender with a browser front end, with working code in six languages.

​
Learn more

​
What’s next

Next, you’ll learn how token pricing works, how prompt caching lowers it, and how to pick the right model and reasoning effort for each job.

Lesson 8: Cost and Speed

Token pricing, prompt caching, choosing a model and reasoning effort.