fair-rose•15mo ago
Knowledge base has problems with misspelling
Hello!
A client of mine recently complained that the knowledge base AI has problems with typos. He will show me an example next week, but in my tests I also found it sometimes to be the case.
To be more specific:
I downloaded the FAQs of a gym from a website into a file. The language is German.
When I ask how much a trainig unit is, I get the correct answer from the KB.
But: When I switch two letters of the word "training unit" (German correct word: "Trainigsstunde", wrong version: "Trainigsstuned"), then I get no info. This is strange and shouldn't happen.
I also should say that it usually handles misspellings well, but sometimes surprisingly badly. Any ideas what I can do here? ChatGPT has no problems with handling the exact same request
7 Replies
Hello!
It sounds like you're experiencing issues with the Knowledge Base (KB) handling misspellings in queries. Here are some steps and tips to improve the robustness of your KB in handling such cases:
Tips for Effective Previewing:
1. Test with Variety: Use a wide range of queries, including those phrased differently, to ensure robustness in the AI’s responses.
2. Check for Consistency: Ensure that the responses are consistent with previous answers and with the data provided.
3. Test with Different Settings: View the previewing process as iterative. The goal is to continuously enhance the AI’s ability to respond accurately.
- Preview-Specific Settings: Changes made in Preview are temporary and will not impact the KB's overall settings.
- Overriding Defaults: To test different behaviors, you can adjust the AI model, temperature, max tokens, instructions, and system prompt which will override the global KB settings.
Permanent Changes:
If satisfied with the preview results, you must manually update these preferences in the KB Settings to apply them globally.
For more detailed guidance, you can refer to the following articles:
- How the Knowledge Base Works
- Response AI
These steps should help improve the KB's ability to handle a variety of queries and provide more accurate responses.
fascinating-indigo•15mo ago
What is the temperature on your model?
fair-roseOP•15mo ago
I tried everything up to 0.8, but same result
fascinating-indigo•15mo ago
Have you also tried different models?
And just to clarify this issue is in KB preview?
fair-roseOP•15mo ago
Yes, I test it in the preview and play around with the parameters. I also tried the available Claude versions but its the same. As it's a free acount, I only have GPT 3.5 Turbo
fascinating-indigo•15mo ago
Yes, sorry I can't help any further. Might be a language issue?
the model is not the issue.