Where could I find more resources to understand about Logica + AI?

[jagandecapri]

Hi,

I came across Logica and it is very interesting! I read about Logica + AI at https://logica-web.github.io/logica-ai/. Where can I read further on the Logica + AI projects? I'm interested to know how Logica can work with LLMs and NeSy seems to be working in the direction.

I've done a search using the keywords "NeSy" and "AI" in this repo's discussion forum and not discussion turned up. Hence, I'm asking this question.

[yilinxia]

Hi,

Thanks for the question!!

The NeSy component of Logica+AI is still under development, and we are continuing to refine it with the goal of an official release.

I recommend trying it out to explore how different aspects of reasoning can be distinguished. For instance, some reasoning can be handled using strict rules in purely Logica, while others are better managed with large language models, like the example we provided https://logica-web.github.io/logica-ai/nesy.html.

You may also be interested in our recent publications on combining LLMs with Logica:

• Natural Language to Logica: Towards Interactive and Explainable Data Analytics
• LogicLM: Robust Application of Large Language Models with Logic Programming for Data Analytics

[yilinxia]

Great question!The program executes only for Bob and Charlie because the EligibleTrip(employee, destination, cost) rule has already excluded Alice - her trip cost exceeds 2000.

[jagandecapri]

Thank you for the helpful answer. It appears to me that enforcement of rules such as EligibleTrip(employee, destination, cost) could be done programmatically, e.g., using Python, or through SQL and then, prompting the LLM using the filtered result. In this context, what is the advantage of Neuro-symbolic AI using Logica?

[yilinxia]

@jagandecapri You’ve nailed it! While something like EligibleTrip(employee, destination, cost) could certainly be implemented with SQL or Python, those approaches typically hard-code logic - they can tell you what passes a rule, but not why or how that reasoning interacts with other predicates.

To me, this really highlights the power of formal methods or symbolic reasoning - they enable deeper reasoning about both the rules and the reasoning process itself. With the Intelligence predicate, you can clearly distinguish which parts are governed by strict rules and which are more flexible.

The key advantage of using Neuro-symbolic AI with Logica lies in how it unifies symbolic reasoning (logical rules, constraints, and explainability) with neural learning (e.g., generalization capabilities provided by LLMs).

[jagandecapri]

@yilinxia Thank you. My difficulty is in understanding the value of Neuro-symbolic AI done with tools like Logica. The example in https://logica-web.github.io/logica-ai/nesy.html, as follows, seems to be just filtering the data before feeding it to the LLM.

# Combine strict and flexible reasoning
TripRequest(employee, destination, cost, visa:visa) :-
  Trip(employee:employee, destination:destination, cost:cost),
  Employee(employee, nationality:nationality),
  visa = Intelligence(NeedVisaAssessment(nationality, destination)),
  EligibleTrip(employee, destination, cost);

I was thinking along the lines of how I can pitch to my team on AI approaches like Neuro-symbolic AI with Logica. At the moment, I could see the value in readability and maintainability of the code, as the logic is not hard-coded in a different language like Python / SQL. But, I can't see the advantage in terms of functionalities that is only possible in Neuro-symbolic AI with Logica compared to the conventional way, e.g., Python / SQL + LLM. This is where my difficulty lies.

[yilinxia]

@jagandecapri That’s very true, and I DO think we need showcase more specific scenaiors that really highlight Logica’s strengths. We would appreciate if you could join the effort : )

In terms of nenurosymbolic AI, if you want robust, sound, and verifiable formal reasoning in nontrivial domains, you can’t rely purely on LLM + Python/SQL. You generally need a symbolic or logic component (or a hybrid/neurosymbolic architecture) to guarantee correctness, that makes Logica critical.