Crop advisory rules extraction and chat setup - all a weekend's play!

[Aaditeshwar Seth]

Hi all, 

We've been working with the Water Data Exchange consortium, an upcoming framework to make water related data and APIs more easily accessible and interoperable. As a PoC, the group decided to build a crop advisory use-case, that given a location and crop and sowing date, can we get weather forecast information and consult a rules-table to send back an advisory for a farmer. The Sattva team came up with a simple rules-table which organizing the crop growth into stages, and for each stage gave rules of what to do if there is light rainfall, or moderate rainfall, heavy rainfall, etc. This is what a JSON equivalent of an Excel file looks like for a sample crop. 

I was curious if we could automatically build something similar for other crops. I found these advisory bulletins by ICAR and wanted to see how far we can go with parsing them into structured formats like the rules-table above. The Cursor IDE was my companion and we managed to build this in a little over a day! Try it out. 

https://act4d.iitd.ac.in/agromet-advisory/

We first extracted text from the PDFs, then used LLMs to generate region-wise JSONs for each crop. Since the bulletins were time-based, this was something like converting from a column format to a row format, to get crop-wise JSONs. There was lots of cleaning to do after that, like to merge JSONs for paddy and rice (same crops), handle overlapping time windows, rotate to fit agricultural seasons, etc., and also an important cross-verification using LLMs to double check the final JSONs against the original bulletins and flag any disputed cases. 

Finally we built a chatbot like frontend for users to ask questions, and also for users to give feedback on the advisories. And a moderation interface for experts to review the advisories, suggestions, and make edits to the advisories. These edits automatically go to git, and can be pushed from there to the rules repository on github. You can check out the review page using login/password = demo/demo!!demo

All in all, the conclusion: The advisories extracted from the bulletins are not great or even very correct, and making them truly useful for a farmer will need a lot more context perhaps specific all the way to the field of the farmer, but to me this tells two things: (a) there is a lot of spoke to extract structured information from text sources, and (b) being able to describe more and more of the world in data can help make knowledge transferrable. No idea where we'll end up with the kind of AI tools we have now that can build all this stuff so rapidly!

https://github.com/aaditeshwar/agromet-advisory

Adi

--

Aaditeshwar Seth
Microsoft Chair Professor, Computer Science and Engineering, IIT Delhi

Co-founder, Gram Vaani; Co-founder, CoRE Stack

Research | CoRE Stack | Gram Vaani | ACT4D 

Technology and (Dis)Empowerment: A Call to Technologists

Latest@GV: Facebook | Twitter | LinkedIn

[Prabhakar Rajagopal]

Hi, 

Tried the crop advisory. Not very useful. No way to better the advisory with more user details... 

Prabhakar

--
You received this message because you are subscribed to the Google Groups "CoRE stack: NRM" group.
To unsubscribe from this group and stop receiving emails from it, send an email to core-stack-nr...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/core-stack-nrm/CAAQC%2Bc-MkfsChvD87GaEQFtfmaiiYt%2B7MPLaRyjQzYy0UMhJFA%40mail.gmail.com.
For more options, visit https://groups.google.com/d/optout.

[Aaditeshwar Seth]

Totally agree, Prabhakar! This is merely a PoC of what can be done using context of geography (state/region level only) and timing (although this is very sparsely populated for most crops, E.g. only one of two weeks of information). Crop varieties, field conditions, past history, etc. nuances can't be handled because none of that info was available in the docs used. To make it truly useful we need to find resources which have this information, but even then this kind of advisory would be of limited value I feel -- farmers may want very specific details like the exact amount of fertilizer to add but this will be extremely difficult to get right, or farmers may want suggestions on how to spot for certain pests and which will require tying this up with more detailed pest information (possible to do, website likes apnikheti and pestoscope have this information and can be crawled and indexed). 

regards
Aadi

To view this discussion visit https://groups.google.com/d/msgid/core-stack-nrm/CABzdB7YPF3WhAFUw6-ZfwFwEc94kOsJuUGUb5EA1VqxCn356qw%40mail.gmail.com.

For more options, visit https://groups.google.com/d/optout.

CAUTION: This email originated from outside of IIT Delhi. Do not click links or open attachments unless you recognize the sender and know the content is safe.

--
Aaditeshwar Seth
Microsoft Chair Professor, Computer Science and Engineering, IIT Delhi

[faiz alam]

Hi Aadi

See this  SukhaRakshak AI (this integrates forecast + advisory - chatbot for farmers, extension agents..) some colleagues here have been working on this with ICAR. From my limited understanding, the bottleneck is a lack of updated and context-specific advisories, with the current district contingency plans being old, also don't capture variations within the district. I think ICAR was working on updating the same for pilot blocks

regards

Faiz

To view this discussion visit https://groups.google.com/d/msgid/core-stack-nrm/5d0f38e3-d719-4126-8e50-2515acda7d7c%40cse.iitd.ac.in.

For more options, visit https://groups.google.com/d/optout.

--

Mohammad Faiz Alam

B.Tech Civil Engineering

IIT-Delhi

[Aaditeshwar Seth]

Great, this is excellent! The advisory is not restricted just to drought protection, it also gives specific info on pest attacks, and the question suggestions at different stages are very helpful. Will be great to see what parts people find useful. Clearly the scope is great that if you are a domain expert you can now easily build such systems, or provide inputs to improve the reliability of these systems. 

regards
Aadi