LLMTransform component - Base/Custom URI for other providers

[jpaubzant]

Hello, I'm not able to connect to LLMTransform component with my custom API url.

Describe the feature request

I would like to request the addition of a "Custom" provider option in the LLMTransform component. Currently, I am unable to connect to my custom/self-hosted LLM API because the component only supports predefined providers.

Having a "Custom" provider would allow users to connect to any OpenAI-compatible API (like vLLM, LiteLLM, or private proxy gateways) by configuring their own endpoints and headers.

Proposed Solution: When selecting "Custom" as the provider, the component should expose the following input fields:

Base URL: (e.g., https://my-custom-gateway.internal/v1)
Endpoint/Path: (e.g., /chat/completions or /completions if not standard)
Custom Headers: A key-value pair input (useful for custom authentication, e.g., X-Gateway-Token, Organization, etc.)
API Key: (Standard field, but mapped to the standard Authorization: Bearer header)

[SouravRoy-ETL]

On it!

[SouravRoy-ETL]

Shipped on main (commit 6dea12d). The AI transforms (LLM, Classify, Embeddings) now have a "Custom (OpenAI-compatible)" provider plus three optional fields:

• Base URL, e.g. https://my-gateway.internal/v1
• Endpoint path, e.g. /custom/chat/completions (blank uses the standard /v1/chat/completions or /v1/embeddings)
• Custom headers, key/value pairs for gateway auth (X-Gateway-Token, Organization, etc.)

The API key still maps to Authorization: Bearer, and a custom Authorization header takes precedence over it. The endpoint path joins onto the Base URL, so a base ending in /v1 will not double up. Pipelines with no overrides keep hitting the standard OpenAI paths unchanged.

Verified end to end against a local OpenAI-compatible mock: the request hit /v1/custom/chat/completions with both the Bearer key and the custom X-Gateway-Token header. It will be in the next release.

[SouravRoy-ETL]

v0.5.4 is out. The AI transforms (xf.ai.llm, xf.ai.classify, xf.ai.embed) gain a Custom provider with an optional base URL, endpoint path, and extra headers, so you can point them at any OpenAI-compatible endpoint.

Grab it here: https://github.com/slothflowlabs/duckle/releases/tag/v0.5.4 (or use the in-app updater). Leaving this open until you can confirm the fix on your setup - let us know how it goes.

[jpaubzant]

Helllo, I'm testing the new feature and connection is working fine, however it seems that "Prompt template" parameter is not mapped with the content key from openai ai swagger, it always sending request with empty content. From the log from my ai endpoint, always having 'content : ''.'

Thx a lot for your work

[SouravRoy-ETL]

Good catch, and thanks for the clear symptom (content: ''). Fixed on main (commit 78c2283).

The prompt template never reached the request. The properties panel wrote the prompt under one key while the engine read it under a different one, so the template came through empty and every call sent an empty user message. That is now aligned, and the engine also accepts the old key, so a pipeline you already saved keeps working without re-entering the prompt.

One related detail: the template placeholder syntax is single braces, {column}, not {{column}}. The form hint was showing the double-brace form by mistake, which is also corrected.

I re-ran it end to end against a local OpenAI-compatible endpoint: with a Say hi to {name} template over a row where name = Alice, the request now sends content: "Say hi to Alice" instead of empty.

While in here I found and fixed the same key mismatch in the sibling AI transforms (Classify, Embed, Chunk), so those were operating on the wrong column too and now honor the column you pick.

On main now and will be in the next build. Leaving this open so you can confirm once it is available.