Roadmap: Assisting LLMs with Kedro knowledge and user consumption #5130
Description
Problem Statement
Many users already rely on GenAI tools to help them build with Kedro — from starting new projects, to learning about Kedro concepts, to building pipelines. Vibecoding is already happening.
However, LLMs alone are not always able to provide users with accurate or up-to-date information. This leads to friction in the user journey:
- Incorrect CLI commands (e.g., suggesting the obsolete --package flag in kedro new)
- Wrong dataset types (e.g., defaulting everything to Parquet instead of using MatplotlibDataset or PlotlyDataset)
- Other subtle errors that slow down learning and adoption
We’ve also observed that prompt quality has a huge impact: LLMs perform much better when guided with clear, structured prompts. Without that, users often get partial or misleading workflows.
Description
We want to explore how Kedro knowledge can be effectively consumed by LLMs and surfaced to users in practical ways.
Currently, LLMs produce outputs that are partially correct but not fully reliable. This exploratory work will focus on:
- Developing prompts/instructions that improve LLM reliability for high-value Kedro workflows
- Exploring delivery mechanisms for these prompts beyond MCP, such as:
- A prompt library in the Kedro docs
- Prompts/instructions embedded directly into kedro new or starter templates.
- Defining approaches for benchmarking LLM outputs so we can measure improvement over time.
What we already know
We conducted interviews with several major Kedro users (large clients with broad adoption) and identified high-value workflows where LLM assistance could provide meaningful benefits. These workflows were also highlighted by users themselves, including:
- Converting notebooks into Kedro projects
- Migrating projects to Kedro 1.0
By refining prompt design, embedding guidance directly into user-facing tools, and systematically measuring performance, we can make LLM-powered Kedro usage both more accurate and more practical.
Approach
We are focusing our initial exploration on two minimum viable prototypes (MVPs) that emerged as high-value workflows during user interviews:
- Converting notebooks into Kedro projects
- Migrating projects to Kedro 1.0
Ask for the Community
We’d love your input and involvement in shaping these experiments:
- Testing: Would you be interested in trying out these prototypes once they’re available?
- Ideas: Are there other Kedro workflows where LLM-powered assistance could add real value?
Please share your thoughts directly in the comments below or on the linked tickets. Your feedback will directly guide how we prioritise and refine this work.
What key metric is this targeting?
- User adoption
- an increasing number of people are using vibe coding in their work, and this could allow Kedro to be used by a wider audience. It will also potentially unblock existing users, such as Kedro users who want to convert Jupyter notebooks to Kedro but who don't have the time
SPIP Questions
Q1. What are you trying to do? Articulate your objectives using absolutely no jargon.
- Enable LLMs to be more effective at creating Kedro pipelines, ideally from within their chosen IDE (eg. Cursor)
Q2. What problem is this proposal NOT designed to solve?
- Not fine tuning LLMs
Q3. How is it done today, and what are the limits of current practice?
- Currently LLMs make a good attempt with Kedro, but there are often blockers such as using the wrong version of the code, or struggling with more complex end-to-end tasks like Jupyter notebook conversion
Q4. What is new in your approach and why do you think it will be successful?
- Augmenting the existing capabilities of LLMs is likely to be successful, because as LLMs get better then the solution will also get better
Q5. Who cares? If you are successful, what difference will it make?
- Any user who uses LLMs to create Kedro code will benefit. This is likely to be an increasing number of users over time. For some of these users it will be the difference between using or not using Kedro
Q6. What are the risks?
- Any LLM related work will need guardrails. There is a risk that the functionality will soon be redundant, given the rapid pace of genAI developments
Q7. How long will it take?
- Weeks/Months
Q8. What are the mid-term and final “exams” to check for success?
- LLMs stop making version related mistakes
- Users can convert Jupyter notebooks to Kedro pipelines