Connect AI models to your data without leaking PII

Let your AI features answer from your private data, with each answer limited to the files and records that the asking user can open. Before a prompt reaches a third-party model provider, Conjoin AI redacts the personal data in it.

Get early accessGet in Touch

AI answers from your private data, scoped to each user's permissions

Conjoin AI indexes the files in your Storage containers and the records in your Database tables so that your application can search them in plain language when it builds a prompt. When you send one prompt to several models in parallel, Conjoin can also merge their answers into one response.

Summarize reports and stream answers

Agent is ready

Explore AI capabilities

Each card links to a page that covers one part of Conjoin AI in more detail.

Context

Context returns results only from the sources that the user who sends the query can open.

Explore

Policy

Policy redacts sensitive data from every prompt and stops AI spending at the budget that you set.

Explore

Inference

Inference sends each request to the model that you name, so you can switch models with one parameter.

Explore

Workflows

Write each workflow in JSON as a trigger, the conditions to check, and the AI steps to take.

Explore

What Conjoin AI adds to your application

Identity-aware context

When a context query includes a user ID, Conjoin limits the results to passages from sources that the user's account has access to. That access comes from the Storage and Database access rules that your application already uses, so you can maintain one set of rules for both the application and its AI features.

Policy firewall

The policy firewall inspects each prompt before Conjoin sends it to a model, and each response before Conjoin returns it to your application. A rule that matches a card number, a social security number, or another kind of personal data can redact the match or block the request.

Multi-model inference

In one request, Conjoin calls several models with the same prompt concurrently and returns every model's response for you to compare. With synthesize: true, Conjoin passes those responses to a judge model, which merges them into a single answer.

Declarative workflows

A workflow is a JSON list of AI steps. A step can also call another Conjoin service, such as Messaging, which can send an SMS with the text that an earlier AI step produces.

Zero-ETL indexing

When you turn on indexing for a Storage container, Conjoin parses each file and stores its embeddings in an index. Conjoin also hosts the embedding pipeline and the vector database for that index, so the indexing toggle is the only setting that you change.

Budget controls

Set a spending limit for a project, a team, or a single user. Conjoin rejects every further AI request once spending reaches that limit.

Multi-modal generation and understanding

To analyze an image, an audio file, or a video, pass its Storage path as the input. When you generate an image or a video, Conjoin writes the file to the Storage path that you name as the output. Conjoin checks the requesting user's Storage permissions on both paths, so a user can process and create files only where their account has access.

Observability

Conjoin records a trace for every AI request, with its token counts, a latency breakdown, and its cost in US dollars. Group those traces by project, team, or user to compare spending, or open one trace to find the step that adds the most latency to a response.

0
ETL pipelines

Pipelines that you build to make Storage and Database searchable

100%
Permission sync

AI queries that Conjoin scopes to the user's existing access

4x
Multi-sample

Samples of one prompt from a single model

Build faster with solutions designed for real-world needs

Every AI query returns only what the user can already open

Pass the requesting user's account ID as the user ID on each context query. Conjoin checks every matching passage against that account's access in Storage and Database before it adds the passage to the results.

Impact

The sample requests up to eight passages about the Mossridge AI contract, which Conjoin selects only from documents within the sample account's Storage permissions.

Prompt for a coding agent: Find passages about the Mossridge contract, using only documents that this user can open.

Conjoin IDE

Select a file to view its contents

Redact PII before Conjoin sends a prompt to an external model

The policy check takes the outbound prompt and returns whether Conjoin allows it, a copy with card numbers and other personal data redacted, and the list of rules that match. Conjoin stores the original text in your request logs, while the provider receives only the redacted copy.

Impact

The firewall replaces the card number in the declined-card ticket for Topiary Health before your code sends the prompt to any model. When the firewall blocks the prompt outright, the sample throws an error before any model call.

Prompt for a coding agent: Redact card numbers from each support ticket before we send its prompt to any model.

Conjoin IDE

Select a file to view its contents

Compare several models and merge their answers into one

Ask several models the same question in parallel, or request up to four samples of one answer from a single model. If you also turn on synthesis, the judge model that you name in the request compares the responses and writes the final answer.

Impact

Conjoin returns one radiology recommendation for Crossbeam Therapeutics that combines the responses of every model in the request, along with the request ID, cost, and latency that the sample saves for an audit.

Prompt for a coding agent: Ask several models for a radiology recommendation, and merge their answers into one.

Conjoin IDE

Select a file to view its contents

Chain AI steps with other Conjoin services in one JSON workflow

A workflow can call other Conjoin services between its AI steps. When your support system sends a webhook for a new ticket, for example, one workflow can summarize the ticket with an AI model, look up the customer in Database, and send the summary to the on-call support lead through Messaging.

Impact

Conjoin passes the data between those services inside the workflow, so your support system makes one webhook call for the whole chain.

Analyze and generate media through Storage paths

A claims app can send a vision model the Storage path of a photo that a customer uploads and receive a description of the damage. The same app can generate an annotated image and have Conjoin write it to the claim's folder in Storage, where the claim's existing access rules apply to the new file.

Impact

Your code passes Storage paths in both directions, so Conjoin moves the files between Storage and the model for you.

Common questions about Conjoin AI

Build your next app on Conjoin

Conjoin will open with a free Build tier, while the pricing page already lists the launch prices for Pro and Business. If you tell us what you're building, we'll talk you through early access.

See pricingGet early access