Spike: Quick win GenAI features

[merelcht]

Goal

Look into quick wins we can add to Kedro to make it easier to build GenAI projects.

Some starting points:

• Vector DB (https://github.com/kedro-org/kedro-academy/blob/main/kedro-rag-chatbot/src/kedro_rag_chatbot/datasets/vector_store_dataset.py)
• GenAI starter templates
• prompt datasets
• agent trace logging
• deeper integration with LLM evaluation frameworks
• pre-built Kedro pipelines that can be used in any GenAI project (e.g. Data processing)

Deliverable

• Scoped & estimated tasks for min. of three quick win GenAI features

Size: S

[deepyaman]

• I haven't looked into it much, but https://github.com/apache/burr seems like it could be a relevant product in a similar space to look at.
• @ianwhale did some early work in this space IIRC (e.g. the LangChain datasets).

[ianwhale]

• I haven't looked into it much, but https://github.com/apache/burr seems like it could be a relevant product in a similar space to look at.
• @ianwhale did some early work in this space IIRC (e.g. the LangChain datasets).

Reporting in!

If you're starting to look for evaluation frameworks, I've seen people like opik: https://github.com/comet-ml/opik

[lrcouto]

GenAI starter that we made during the hackaton in London: https://github.com/SajidAlamQB/kedro-genai-starter

Maybe we could take some ideas from this as well.

[ravi-kumar-pilla]

If you're starting to look for evaluation frameworks, I've seen people like opik: https://github.com/comet-ml/opik

Thank you @ianwhale for the pointer. Do you know if anyone or any team tried https://github.com/confident-ai/deepeval