Find answers from the community

Updated 3 months ago

Hello friends I am also new to

Hello friends! I am also new to LlamaIndex and, to try it out, I wanted to test it in a very simple project: I want to create a chatbox that will answer questions using the GitLabs handbook as its context, found here: https://about.gitlab.com/handbook/ . In the project files, I have cloned the repository (found here: https://gitlab.com/gitlab-com/www-gitlab-com/-/tree/master/sites/handbook/source/handbook ) so the files are local, and they are all of ".html.md" extension.

The handbook has more than 2,000 pages, and the index.json that is created is, accordingly, more than 1GB in size. So far so good (or so it seems), but then the issues come up:

1) The chatbox answers don't really relate to the handbook at all. They seem pretty random and they don't seem to have the context. If I ask it, for example, "What are the six principles?" it will give me some six that are not in the handbook. If I ask it "What are the six principles of the company?" it will tell me it doesn't know what company I'm taking about.

2) It will still answer random questions (say, "what is the capital of Bulgaria?") instead of staying in focus.
Attachment
Screenshot_2023-04-19_at_9.55.42_AM.png
V
L
m
12 comments
On the other hand, if I use a Markdown reader, the index.js file that is created is tiny... the text I find inside contains only the "index.html.md" file. However, the chatbox can answer some questions correctly which are not found in index.html.ms (for example, the aforementioned "what are the company's values?" is answered correctly), but not a whole lot.
Attachment
image.png
Thank you everyone for any help or pointer you might be able to give πŸ™‚
Two initial notes:
  1. I would lower the temperature to 0
  2. Make sure you pass the service context back in when loading from disk
  3. Ok I lied, three notes. You might want to increase similarity_top_k in your query, something like this maybe index.query(..., similarity_top_k=3, response_mode="compact")
@Logan M Thanks! I'll try it out later when I can get back to that project and let you know how it goes.
Hi @Logan M thanks for te help
I have the same issue (after upgrading from 0.4 > 0.5.18 )
It looks like the response did not come from the right context
the response takes a lot of time (almost 1 minute), I have 1 doc text file with 10 questions (in V 4.0 everything was good ) can you advise?
@Vayu can you update here after your checj the Logan advise?
I will. I'm just at my job right now. This is a side thing, so I need to make time for it probably later tonight πŸ™‚
Add a reply
Sign up and join the conversation on Discord