First time user of the community platform

[steff_horemans]

Hi everyone,
Don't know where to put this specific question. I'm working on a reference data mesh implementation to connect and combine datasets to find matching trials for patients with a specific genetic profile. 
- Do you know anyone that might be interested in this project? If yes, please show the post and send them to me. 
- I'm still figuring out a lot of things. I want to use GenAI to extract eligibility criteria from structured trial text. Has anyone done this before? If so, what approach did you use: use the ai_queries, API calls to foundation models or download a bespoke model?
- From a governance perspective, how would you store and share a bespoke LLM model? Do you share via MLflow and serving like your own models or is there a way to extend Foundation API's with models of your own organisation for discovery?

https://www.linkedin.com/posts/steff-horemans-34884497_raredisease-dataengineering-precisionmedicine...

[Ashwin_DSA · Accepted solution]

Hi @steff_horemans,

Yes, this is absolutely fine to ask here.

You're touching a few quite different areas, though... Trial matching/reference data design, GenAI extraction of eligibility criteria, and governance/serving of bespoke models. You’ll likely get better responses if you either:

• Split this into a couple of more targeted questions, or
• Add a bit more detail on your current architecture, data format, and where you are blocked.

For example, it may help to clarify things like:

• Whether the trial text is already structured or only partially structured
• Whether you want extraction, classification, or matching
• Expected scale and latency
• Governance constraints around PHI / sensitive data
• Whether you are deciding between Foundation Model APIs, SQL/AI functions, or a custom fine-tuned workflow

You may also get stronger engagement by posting the most relevant part in a focused board, such as generative AI, Machine Learning, Data Governance, or Data Engineering, depending on which question you want answered first. That way, people will be able to give much more concrete advice.

And now to your questions... at a high level, my instinct would be..

• Start with prompt-based structured extraction first, using ai_query or model API calls, since Databricks supports querying foundation, external, and custom model endpoints from SQL, and there are community examples of using AI_QUERY for structured extraction before moving to fine-tuning.
• Only move to a bespoke fine-tuned model if prompt-based approaches are not accurate or consistent enough for your use case.
• Treat a bespoke model as a governed model asset, for example by packaging it in MLflow, registering it in Unity Catalog or the workspace registry, and serving it through Model Serving.
• Assume custom organisational models would normally be shared via your own serving endpoints rather than added directly into Foundation Model APIs, which are for Databricks-hosted foundation models. Externally hosted models are handled through external model endpoints.

Hope this gives a direction.

If this answer resolves your question, could you mark it as “Accept as Solution”? That helps other users quickly find the correct fix.