📣
TiDB Cloud Premium is now in public preview. Unlimited growth, instant elasticity, advanced security for enterprise workloads. Try it out →

Configure TiDB Cloud Filesystem AI Providers



A TiDB Cloud Filesystem can optionally extract text from media files and generate embeddings for stored content. To enable these capabilities, you can configure one or more AI providers through the CLI.

Prerequisites

The configuration commands require TiDB Cloud API credentials and an explicit Filesystem ID. Set the provider key through TI_FS_AI_PROVIDER_API_KEY. The CLI does not persist the key locally. The Filesystem service stores it encrypted and returns only a masked value in subsequent configuration output.

For an interactive shell, read and export the provider key without placing it in shell history:

printf 'Provider API key: ' >&2 read -r -s TI_FS_AI_PROVIDER_API_KEY printf '\n' >&2 export TI_FS_AI_PROVIDER_API_KEY

In CI, inject TI_FS_AI_PROVIDER_API_KEY from a masked secret. Unset the variable after you finish configuring providers.

Inspect media extraction configuration

Read the effective extraction configuration for a media type:

ti fs describe-file-system-extract-configuration \ --file-system-id "<file-system-id>" \ --media-type image

Update media extraction configuration

Use update-file-system-extract-configuration to enable, update, or disable image, audio, or video extraction. For example, configure image extraction through an OpenAI-compatible provider:

ti fs update-file-system-extract-configuration \ --file-system-id "<file-system-id>" \ --media-type image \ --enabled true \ --provider-api-base https://api.openai.com/v1 \ --provider-model "<vision-model>" \ --provider-protocol openai

The openai protocol supports image, audio, and video extraction. The qwen-asr protocol is supported only for audio extraction through Alibaba Cloud Model Studio. An endpoint from another provider might work if it implements the required OpenAI-compatible API contract. Native interfaces for Anthropic, Gemini, Vertex AI, Amazon Bedrock, and Azure OpenAI are not supported.

To disable extraction for a media type:

ti fs update-file-system-extract-configuration \ --file-system-id "<file-system-id>" \ --media-type image \ --enabled false

Inspect embedding configuration

Read whether embeddings are managed by the application or database:

ti fs describe-file-system-embedding-configuration \ --file-system-id "<file-system-id>"

Update embedding configuration

Use update-file-system-embedding-configuration to update the optional application-managed embedding configuration. For example:

ti fs update-file-system-embedding-configuration \ --file-system-id "<file-system-id>" \ --enabled true \ --provider-api-base https://api.openai.com/v1 \ --provider-model text-embedding-3-small

Application-managed embeddings require an OpenAI-compatible endpoint that returns 1024-dimensional vectors. They are available for Shared Filesystems and Native Filesystems whose effective embedding mode is fts_only. If a Native Filesystem uses database-managed automatic embeddings, the service rejects this update and reports source=database_auto.

After you finish configuring providers, remove the key from the current shell:

unset TI_FS_AI_PROVIDER_API_KEY

To disable that configuration:

ti fs update-file-system-embedding-configuration \ --file-system-id "<file-system-id>" \ --enabled false

Data flow after configuration

After you enable extraction, the Filesystem service sends media content to the configured extraction provider. It sends the extracted text or descriptions to the configured embedding provider. Choose provider accounts and retention policies appropriate for your data.

If an update fails because of a timeout, lost response, or another ambiguous network error, run the matching describe-file-system-*-configuration command before retrying. The provider validation request might have succeeded and incurred a charge even if the CLI did not receive the response.

What's next

Was this page helpful?