# AI Search data usage FAQ

This FAQ explains how Redocly AI Search uses your data, which models are used, and how the feature complies with enterprise privacy requirements.

## Does AI Search train on our data?

No.
AI Search in Redocly does not use your data to train or fine-tune any AI models.

## How does AI Search work?

AI Search uses **hosted large language models (LLMs)** accessed via:

- **AWS Bedrock**, and
- **Google Cloud Vertex AI (Gemini)** — depending on configuration and ongoing model evaluation.


These services run models in **inference-only mode**.

How a search request works:

1. **Embeddings** are generated from your documentation using an LLM (via AWS Bedrock).
These are stored in **Redocly’s own vector database** for fast retrieval.
2. When a user submits a search query:
  - The query is embedded and matched against the vector database.
  - Relevant text snippets are selected and **re-ranked**.
  - A **system prompt**, an optional **project-specific prompt**, and the selected snippets are sent to an LLM to generate an answer.
3. The LLM generates an answer in real time.
No model training or fine-tuning is performed.


## What data is sent to the model?

For each search request, Redocly sends:

- The **user’s search query**
- A **system prompt** (defined by Redocly)
- An **optional project prompt** (customer-configurable)
- Selected **documentation excerpts** (API specs, markdown, etc.) based on vector search and re-ranking


This data is used only to generate the search response.
It is not used to train or adapt any model.

## Are prompts or responses stored by Redocly or by model providers?

- **AWS Bedrock:** does not use customer data, prompts, or responses to train models ([AWS Bedrock privacy](https://aws.amazon.com/bedrock/amazon-models/privacy/)).
- **Google Gemini:** supports inference-only operation without training on customer content ([Gemini data governance](https://cloud.google.com/gemini/docs/discover/data-governance)).


**Redocly:** does not store prompts or responses for model training, and does not share them with model providers.

## Is any of our data used to improve the AI models?

No.
Your data is not used to train, fine-tune, or otherwise improve any underlying AI models used by AI Search.

## Is this compliant with enterprise data privacy requirements?

Yes.
Because the models are used in **inference-only mode** and your content is not used for model training, AI Search supports common enterprise privacy and compliance expectations.

## Can we configure or disable AI Search?

Yes.
You can configure an **optional project prompt** to adjust the tone or style of responses — or you can choose to **disable AI Search entirely** if desired.

See: [AI Search configuration](/docs/realm/config/search)

## Which models does Redocly use?

Redocly AI Search uses LLMs accessed through:

- **AWS Bedrock**
- **Google Cloud Vertex AI**


We may periodically test and update model selections to optimize quality and performance.
All models are used in inference-only mode and follow the same data usage principles.