Skip to content

Latest commit

 

History

History
89 lines (64 loc) · 5.25 KB

README.md

File metadata and controls

89 lines (64 loc) · 5.25 KB
Raw

llmsearch

Dec 31, 2023

I haven't touched this in a long time, but to my surprise there is still some traffic. If you have an issue or need an upgrade let me know! I'm using a slimmed down version in my current project, Owl, which I may release someday, if anyone cares enough I have some updates I can backport to this fuller version.

V1.1 - Now with image search too! just ask chatGPT for an image or picture. Note: default return is 10. If that takes too long to display, just tell gpt how many images you want. e.g.:
find 3 images of Stanford University
It still actually gets 10 urls, so you can probably just ask it to display 3 more ...

Release 1.0 now here! Release 1.0 is designed as a single-user version of web search for LLMs, with a single global curation file, and no tools for editting it (use your favorite text editor on sites.json)

A simple web search tool / endpoint with Pre and Post search processsing.
Why? LLMs need to be able to access the web efficiently.
non-LLM chatbots might want text access to a curated set of websites

llmsearch features:

  1. Uses an LLM to rewrite the query for better search results.
  2. Uses the google customized search API to collect urls.
  3. Filters and prioritizes urls by local whitelist/unknown/blacklist, and by past history of usefulness
  4. Uses concurrent futures to process urls in parallel.
  5. Preprocesses url returns to extract a minimal piece of candidate text to send to gpt-3.5 for final refinement.
  6. Stops when sufficient results are accumulated (two levels - Quick and Full)
  7. As a plugin or direct call it returns json format including result, domain, url, and credibility (see chatGPT screenshot below).

Four ways to use this:

  1. as a chatGPT plugin (if you are a plugin developer)
  2. as a network endpoint for your application
  3. as a standalone command line search tool 'python3 search_service.py'
  4. directly from within a python application: import search_service, then call search_service.from_gpt(query_string, search_level)

You will need google customized search api key and your google cx.

You will also need an openai api.key.

llmsearch currently uses gpt-3.5-turbo internally, so its reasonably cheap for personal use, usually a few tenths of a cent per top-level query.
I expect to release an update shortly that will enable use of Vicuna, maybe 7B 4bit if possible.
*** update - this is stacked behind the llm server I'm building, cloud resources are too hard to find and cpu inference is too slow ***
*** update2 - even 7B is too slow with a single 3090. Will be moving up to dual 3090, maybe if I lower the number of samples.***

Screen shot of chatGPT session with plugin installed:

plugin

NOTES: I am a 'quick sketch' research programmer. Careful methodical programmers will probably be horrified with the code in this repository.
*** If you are one such and have suggestions/edits, I'd love your contributions! ***
Having said that, I use this code base every day, all day long, as a chatGPT plugin. Pretty much the only failures I have seen are when the interall api calls to gpt-3.5-turbo timeout. There is quite a bit of recovery code around those, they are rare except when openai is swamped.

INSTALLATION:

  1. clone the repository
  2. pip3 install -r requirements.txt
  3. add your openai.api_key either as an evironment var or directly in utilityV2.py
  4. add your google credentials either as environment vars or directly in google_search_concurrent.py

to test, try: python3 search_service.py

you should see:

Yes?

To run as a gptPlugin (assuming you are a plugin developer) run: python3 main.py

Note that you will need to edit openapi.yaml and .well-known/ai-plugin.json, as well as setting the corresponding site info in main.py
This actually starts up a pretty std web endpoint you can actually even call from your browser. lookup openapi.yaml for more on configuration options, I just copy-pasted.
The endpoint returns json with source, url, response, and credibility keys.

That should do it.

Site curation and prioritization

  1. the file sites contains a json formatted list of sites (next to last portion of domain name, e.g. openai, usually).
  2. At the moment there is no api to manage this list, but you can manually edit it. A site not listed in sites.json is considered 'Third Party' (a chatGPT recommended term, not mine).
  3. All whitelist site urls will be launched before any Third-Party site urls.
  4. url launch order is futher prioritized by site_stats.json, a record of the average post-filtering bytes per second the site has delivered in previous queries.
  5. comes by default with unlisted sites ok to search.
    If you want whitelisted only, look in google_search_concurrent.py in the search def for a line like 'if site not in sites or sites[site]==1' and change it to 'if site in sites and sites[site]==1'. I'll add a config file and make that a config option someday...

In deciding what to whitelist or blacklist, you might want to review the site_stats.
An easy way to do this is to run python3 show_site_stats.py
show_site_stats.py accepts one optional argument: [new, all]
'new' shows only stats for sites not listed in sites.json (so you can decide if you want to whitelist them)