what espressif is hidding, or what AI knows and i dont

[chegewara]

For very long time i was very skeptic about AI, especially with all this chatGPT hype, but recently i decided to see what is this.
Last couple months i am very intensive playing with AI, because i found it very interesting and useful tool, but it is just a tool and i have to learn how to use it.

For very long time it was just a play, and by play i really mean playing RPG text games, testing different models etc.
Not so long ago i asked about LLM with esp32 support, because "normal" models are hallucinating a lot. I am still a noob in AI matter, but i learned a bit about it already so i decided to write some app, backend and frontend, which will be better than chatGPT or other public models and i have some progress. Two days very intensive and hard work with chatGPT (irony) and i have something which is really promising. I will share the full application code when it will be more mature.

I asked simple question to chatGPT and to "my" model and here is result:
- chatGPT found me only pdf version of espressif documentation (random version) - chatGPT have access to internet
- my model found me a link to esp32-p4 TRM, which i missed on forum (with my model ive been chatting about parlio on P4 before this last question) - no access to internet per se

[krzychb]

- my model found me a link to esp32-p4 TRM, which i missed on forum (with my model ive been chatting about parlio on P4 before this last question) - no access to internet per se

If the bot doesn’t have internet access, how do you provide knowledge sources to it? Do you use RAG?
If you use RAG, do you ingest forum posts as-is, or do preprocessing (e.g., filtering only resolved topics) to reduce noise?

[chegewara]

- my model found me a link to esp32-p4 TRM, which i missed on forum (with my model ive been chatting about parlio on P4 before this last question) - no access to internet per se

If the bot doesn’t have internet access, how do you provide knowledge sources to it? Do you use RAG?
If you use RAG, do you ingest forum posts as-is, or do preprocessing (e.g., filtering only resolved topics) to reduce noise?

Yes, i am trying to build robust RAG. My first version did scraping of espressif docs and i believe it found TRM in there. And no, i dont plan to scrap forum, to not overuse servers (at least im not that good to do it safe yet).

By lack of internet i meant it is not using any tools to live search like chatgpt does, just early scraping for RAG, which i can do also from local files.

To be honest a week later i have only PoC, but i am trying to do a bigger project, not only for esp32, but all kind of documents RAG system which i can feed with docs, github links, code examples etc. But with 0 python knowledge, well, its not easy even with AI assistance. If i ever make it working i may think about adding live search too.

[Sprite]

My first version did scraping of espressif docs and i believe it found TRM in there. And no, i dont plan to scrap forum, to not overuse servers (at least im not that good to do it safe yet).

Poke me if you want to do that - I can probably tell you what to scrape and what to leave alone and whitelist your scraper if needed.