feat: add Eden AI integration (chat generator + embedders)

[MVS-source]

Related Issues

Supersedes #3517 (closed). Re-opened via the monorepo route, as discussed with the maintainers.

Proposed Changes:

Adds a new integrations/edenai/ package that integrates Eden AI with Haystack.

Eden AI exposes 500+ models from providers such as OpenAI, Anthropic, Google, Mistral and more behind a single, OpenAI-compatible API — EU-hosted and GDPR-aligned, with unified billing.

New components:

• EdenAIChatGenerator — chat completions through Eden AI's OpenAI-compatible endpoint (https://api.edenai.run/v3), using the provider/model naming scheme (e.g. anthropic/claude-sonnet-5).
• EdenAITextEmbedder — embeds a string.
• EdenAIDocumentEmbedder — embeds a list of Documents.

The package follows the standard monorepo layout: pyproject.toml (package edenai-haystack, tag pattern integrations/edenai-v*), pydoc config, README, CHANGELOG, an examples/edenai_rag_pipeline.py, the .github/workflows/edenai.yml CI workflow, a .github/labeler.yml entry and a row in the root README.

How did you test it?

• Unit tests for the chat generator (sync + async), the text embedder and the document embedder.
• Manual live verification against the Eden AI API:
  • Chat — mistral/mistral-small-2603 → HTTP 200.
  • Embeddings — openai/text-embedding-3-small → HTTP 200, 1536-dim vectors.
• The edenai.yml workflow runs the suite. The live tests read EDENAI_API_KEY.

Notes for the reviewer

• CI secret required: the live tests need an EDENAI_API_KEY repo secret. I have a scoped key ready — happy to share it through whatever secure channel you prefer. (Note: GitHub does not expose repo secrets to workflows triggered from a fork, so the live-test job won't have the key on this PR; it can run post-merge or via your usual maintainer trigger.)
• Conventions: package name edenai-haystack, tag pattern integrations/edenai-v*, aligned with the other integrations.
• CLA: signed.

Checklist

• I have read the contributors guidelines and the code of conduct
• I added unit tests and updated the docstrings
• I've used one of the conventional commit types for my PR title: fix:, feat:, build:, chore:, ci:, docs:, style:, refactor:, perf:, test:.

[CLAassistant]

All committers have signed the CLA.

[sjrl]

Before we review the implementation, could you clarify what changed since the previous PR? In the earlier discussion, the agreed direction was for Eden AI to host and maintain edenai-haystack in the Eden AI GitHub organization and publish it independently, rather than adding it to haystack-core-integrations as mentioned in this comment #3517 (comment)

Is that still an option you would be open to? Given that Eden AI would own the integration, credentials, releases, and ongoing compatibility with the model catalog, hosting the code in your own organization would likely give you more control over maintenance and release cadence. We could then look at listing it in the official Haystack integrations catalog.

Update: Aligned on slack

cc @julian-risch

[julian-risch]

Is that still an option you would be open to? Given that Eden AI would own the integration, credentials, releases, and ongoing compatibility with the model catalog, hosting the code in your own organization would likely give you more control over maintenance and release cadence. We could then look at listing it in the official Haystack integrations catalog.

@sjrl @bilgeyucel aligned with Eden AI via slack on re-opening their PR. I will set up an API key with their help.

[sjrl]

@MVS-source thanks so much for the contribution! It looks like the commits were made with the identity MVS-source, which isn’t currently connected to a GitHub account.

Could you please add the email address used for those commits to the GitHub account that opened the PR, then sign the CLA with that account? Once that’s done, you can use the recheck link above to refresh the status.

Thanks again for helping us get this sorted!

[MVS-source]

Thanks a lot for the detailed review, @sjrl — really appreciate you taking the time to go through it so thoroughly. 🙏

Just pushed a commit addressing everything:

• dropped the haystack-ai upper pin — verified the suite against both Haystack 2.30 and 3.0
• switched to ToolsType from haystack.tools and added a to_dict to EdenAIChatGenerator (drops api_base_url/organization so it round-trips)
• made the embedder init params keyword-only and let the base class hold timeout/max_retries
• trimmed the tests down to Eden AI-specific behaviour and dropped the async test file
• removed the CHANGELOG and slimmed the README to match the Mistral one

Happy to follow up with the catalog page in haystack-integrations as a separate PR. Re-requesting your review — thanks again!

[sjrl]

@MVS-source thanks for responding so quickly! And for addressing all of my comments.

It should be almost good to go. One thing I did forget to mention is could you add an entry to

haystack-core-integrations/.github/workflows/CI_coverage_comment.yml

Line 27 in f7547d8

- "Test / e2b"

its what we use to track code coverage on integrations.

I'll also give another look over of the code and would definitely appreciate a follow up PR in haystack-integrations

[MVS-source]

Thanks @sjrl! Pushed a commit addressing the chat generator comments:

• added the @component decorator
• switched to the explicit super(EdenAIChatGenerator, self) form since @component recreates the class
• re-implemented to_dict with default_to_dict following the cometapi pattern, so only the __init__ parameters are serialized (api_base_url and organization are dropped) and it round-trips regardless of upstream OpenAIChatGenerator changes

On SUPPORTED_MODELS: I matched the sibling chat generators (cometapi, mistral, etc.), which don't define one on the chat class, since Eden AI is a 500+ model gateway. Happy to add a non-exhaustive list for consistency with the embedders if you'd prefer.

Validated the full suite against Haystack 2.30 and 3.0. Thanks again!

[sjrl]

Thanks @MVS-source!

For this point

On SUPPORTED_MODELS: I matched the sibling chat generators (cometapi, mistral, etc.), which don't define one on the chat class, since Eden AI is a 500+ model gateway. Happy to add a non-exhaustive list for consistency with the embedders if you'd prefer.

no worries then, that is a lot of models 😅

[sjrl]

@MVS-source one last thing, could you do what's described in this comment? #3615 (comment)

[MVS-source]

Done, added - "Test / edenai" to .github/workflows/CI_coverage_comment.yml (between e2b and elasticsearch).

And yes, I'll follow up with the catalog page PR in haystack-integrations. Thanks!

[sjrl]

Thanks for the contribution!