AI-assisted extraction of architectural parameters from RISC-V specifications (RISC-V Mentorship)

Mentors: Ajit Dingankar and Allen Baum

Company: Qualcomm Technologies, Inc.

1. Develop prompts & context size for LLMs that identify architectural parameters in the RISC-V         privileged and unprivileged spec.
   a) Give examples from the manually created list (in UDB).
   b) Identify with textual excerpt from the specs.
   c) Classify the parameters as either dependent on a configuration, named or unnamed.
   
2. Use those prompts to collect a list of parameters using at least two different LLMs and the parameters documented in UDB.
   
3.  Refine the prompts to address any differences between LLM and UDB results, as needed.
   a) Filter out any hallucinations as negative examples, e.g. doesn’t fall into any of the 1(c)                          categories or falls into the wrong category (could be context size dependent).
   b) Use some missing parameters as further prompt examples.
   
4. Create a Github PR to insert tags into the spec for each parameter in the resulting list, giving unique names to unnamed parameters according to the tag format spec, and using existing parameter names as examples.
   

Repository URL: https://github.com/riscv-software-src/riscv-unified-db

Learning Objectives:

Open-source collaboration across RVI projects and groups listed below.

1. Machine analysis and understanding of English specs.
2. Prompt engineering for LLMs.
3. Analysis (especially prevention of LLM hallucination) and presentation of results.

Coding Challenge: 

https://docs.google.com/document/d/1IbOHtC-BbKbRd_3a6rB_8158IF9iR7Pi/edit?usp=sharing&ouid=104823481179142905076&rtpof=true&sd=true