3 ms·

I want a simple RAG. Not even the RAG part, just AI backed searching of a document. I put a CSV in there. I ask it a question. It pops me back the closest match

by muzani 2y ago

I want a simple RAG. Not even the RAG part, just AI backed searching of a document. I put a CSV in there. I ask it a question. It pops me back the closest match. I don't need it to be very smart, but it should be better than fuzzy searches. Specifically it has to match a question from a different language, and it has to match questions asked in different contexts. I want users to ask it questions like "I don't like the books recommended here" and it should eventually match up with "What are your favorite books on Indian history?" inside the CSV. So it's not just keywords, semantics matter too. I'm surprised it doesn't exist yet. Plenty of tools do this offline. There's online cloud ones just hosting the vector DB. Not much that just handles a search API. But it has to do this at scale, for tens of thousands of users, not just one user like most sites do. I feel like Algolia should be able to do this easily but it's not on the list.

OrionSeven 2y ago
So does the CSV have multiple languages within it? Why a CSV verse various documents in a directory (like Google Drive?). Just curious, I'm working on AskJack.io and the multiple users, find anything, no matter the context, is exactly what we're doing. But we haven't tried to tackle languages yet.
- muzani 2y ago
  "So does the CSV have multiple languages within it?" For context, it's for Indonesia. So you do have urban areas and expats who can mostly speak English. But English isn't taught in school. Arabic, Chinese, Dutch are popular too. Bahasa Indonesia is the formal version, but most will be more familiar with the various languages. Most states will have their own dialects, at least 5 major dialects. The difference is not like American/English, but closer to Scots/English. This is a representation of the word fruit in different dialects: https://www.facebook.com/photo/?fbid=2078468635856668 https://www.facebook.com/photo/?fbid=2078468635856668 The challenge is that many people are comfortable with writing in their own dialect, even if they can read others. So we can't possibly use all the languages that could be input, but it's fine if it outputs in a somewhat different dialect. For the most part, sentence structures are similar and "bua" and "buah" are still recognizable to people with different dialects, likely to AI as well. But programmatic algorithms wouldn't handle something like "buwa" properly. To answer the question, the CSV will mostly just have informal Bahasa Indonesia. English questions will miss sometimes, so we've added English as well for the ones that miss. More data is more effective. "Why a CSV verse various documents in a directory (like Google Drive?)" Could be anything really, but we tend to have a loose format for data & notes, and then have a script that strips out or combines from other sheets into something that's cleaner to read and search. PDF, XLS, etc is hard to get consistent. But something like JSON, XML, YAML are still acceptable as long as we can script it. AskJack seems to be a bit low code though. We can build our own and have reasons to have more control over certain bits. So a B2B solution over the RAG part would be nice.
floydnoel 2y ago
perhaps https://proompty.com https://proompty.com would meet your needs?
- muzani 2y ago
  This is interesting, thanks! There's some dodgy bits, but I'll check it out.

by muzani 2y ago

OrionSeven 2y ago

So does the CSV have multiple languages within it? Why a CSV verse various documents in a directory (like Google Drive?). Just curious, I'm working on AskJack.io and the multiple users, find anything, no matter the context, is exactly what we're doing. But we haven't tried to tackle languages yet.

muzani 2y ago
"So does the CSV have multiple languages within it?" For context, it's for Indonesia. So you do have urban areas and expats who can mostly speak English. But English isn't taught in school. Arabic, Chinese, Dutch are popular too. Bahasa Indonesia is the formal version, but most will be more familiar with the various languages. Most states will have their own dialects, at least 5 major dialects. The difference is not like American/English, but closer to Scots/English. This is a representation of the word fruit in different dialects: https://www.facebook.com/photo/?fbid=2078468635856668 https://www.facebook.com/photo/?fbid=2078468635856668 The challenge is that many people are comfortable with writing in their own dialect, even if they can read others. So we can't possibly use all the languages that could be input, but it's fine if it outputs in a somewhat different dialect. For the most part, sentence structures are similar and "bua" and "buah" are still recognizable to people with different dialects, likely to AI as well. But programmatic algorithms wouldn't handle something like "buwa" properly. To answer the question, the CSV will mostly just have informal Bahasa Indonesia. English questions will miss sometimes, so we've added English as well for the ones that miss. More data is more effective. "Why a CSV verse various documents in a directory (like Google Drive?)" Could be anything really, but we tend to have a loose format for data & notes, and then have a script that strips out or combines from other sheets into something that's cleaner to read and search. PDF, XLS, etc is hard to get consistent. But something like JSON, XML, YAML are still acceptable as long as we can script it. AskJack seems to be a bit low code though. We can build our own and have reasons to have more control over certain bits. So a B2B solution over the RAG part would be nice.

floydnoel 2y ago

perhaps https://proompty.com https://proompty.com would meet your needs?

muzani 2y ago
This is interesting, thanks! There's some dodgy bits, but I'll check it out.