AI
AI on Ulabase
Your service can search your data by meaning and let AI agents use it, with the AI providers and keys you choose.
AI on Ulabase does two things. Your service can search your documents by meaning, with vectors: a vector is a list of numbers that an AI model computes from a text, and texts with a similar meaning get close vectors. And an AI agent such as Claude can use your data over MCP.
Use it to find the right products in a catalog when the words of the search are not the words of the descriptions, or to answer questions about the files your users upload. A reranking model can then put the best results first.
What you get:
-
Your models, your keys. You choose the providers and call them with your own keys.
-
Vectors with no code. The service computes a vector for every document it stores.
-
Your data stays in your service. The vectors stay in MongoDB, next to the data they describe. Only the calls to your provider leave the service.
The pages of the AI group in the sidebar follow the order you need them: the key, the vectors, the files, the agent.
| Page | What it is for |
|---|---|
The providers the service embeds and reranks with, each with your key and its default models. Everything else in this group runs on what you set here. |
|
A rule per collection: every document written gets a vector for one of its fields, computed by your provider. Under the rule, what is still missing for a semantic search, each with a button to the page that sets it up. |
|
Rules on a file bucket: every file uploaded that matches one is split into text chunks, in a collection that embeds them and that you search like any other. |
|
Let an agent such as Claude use your collections, aggregations, change streams and GraphQL apps, within the permissions you defined. |
From a field to a semantic search
Three steps: a key, a rule, a search.
-
A key under AI Keys lets the service embed text.
-
A rule under Auto Embeddings tells a collection which field to embed. Every document written from then on gets its vector.
-
An aggregation with
$vectorizeand$vectorScanturns a question into a vector and returns the closest documents, with no index. Arerankblock hands the best of them to the reranking model. Published on MCP, the aggregation is a search an agent can run.
When the collection grows, a vector index under Indexes, with $vectorSearch in place of $vectorScan, makes the search fast. Vector indexes and $vectorSearch come with Pro and above.
Auto Embeddings walks you through the three steps on catalog, the collection of the ecommerce starter, so MCP, the search and the shop all use the same products. Each step is a click, and the page also shows the same calls in curl, HTTPie, JavaScript and Python.
What it costs
Your provider bills the calls to your account: one embedding per document written and per question asked, one reranking per search that asks for it. Ulabase adds nothing. Searches run on your service, within your plan. $vectorScan reads every vector it compares, so an index pays off as the collection grows.
Where it applies
The AI pages of the console work on Free, Starter and Pro services. On an Enterprise service we set up semantic search with you, on the MongoDB it runs on: ask us.