Bill Kish

The OpenAI completion API can fail due to various reasons including network errors or more commonly their RateLimitError, or ServiceUnavailableError. This adds a retry() decorator that implements an exponential backoff...

text-davinci-003 supports a context size of 4097 tokens so return 4097 instead of 4000 in modelname_to_contextsize() for text-davinci-003

provide visibility into final prompt

11

For debugging or other traceability purposes it is sometimes useful to see the final prompt text as sent to the completion model. It would be good to have a mechanism...

consider using sqlite + hnsqlib for embedding search/storage

4

very cool concept here. For similar types of projects I have found using sqlite database + hnswlib useful since with pinecone I still typically need another database anyway for persisting...

After building a bunch of different llm apps I found that most of them don't require much more than hnswlib + sqlite for retrieval. This combo scales up to millions...

Exception when number of filtered items in index is less than requested K

6

When using filters and requesting a K larger than the number of filtered items in the index, hnswlib will raise the following exception: "RuntimeError: Cannot return the results in a...

a common failure mode is that we fail to extract the actual story text from a web site for a variety of reasons. (paywall, javascript requirement, anti-bot logic, etc) make...

twitter links produce hallucinations

1

the title text and javescript not available response make the model produce some interesting hallucinations. Should filter twitter links or figure out a way of accessing the content.

currently uses a timeout of 30 seconds with no retries with attempting to retrieve story url content. consider adding mechanism for retry, or enqueue overloaded links for download in the...