(Spelling Issue) - Assistant taking incorrect spel...
# support
s
Even though the prompt explicitly includes this instruction: "Remember the spelling of Accelnomics properly, as sometimes users may give the wrong spelling – you need to take the correct one." The assistant still fails to recognize or use the correct spelling of Accelnomics, which is both strange and unacceptable from a prompt reliability standpoint.These kinds of inconsistencies are concerning, especially when the prompts are very clear and specific. Please investigate thoroughly. You can refer to the following call IDs for your analysis:
ddc6a675-50c2-4a93-84f0-6eb07b4a32a6
ccdc3c3d-304c-4c51-8079-02f9288b7991
v
Hi there! I understand your frustration with the assistant not recognizing the correct spelling of "Accelnomics" despite explicit instructions in the prompt. While I can't analyze those specific call IDs without access to the actual call data, I can definitely help explain potential causes and suggest some approaches to resolve this issue. ## Why This Might Be Happening Several factors could be affecting the prompt reliability: 1. **Speech-to-Text Transcription Issues**: According to the [documentation on multilingual support](https://docs.vapi.ai/customization/multilingual), Vapi uses speech-to-text transcription as part of its process, which can affect how company names are interpreted. If the speech-to-text component incorrectly transcribes "Accelnomics," the assistant won't receive the correct spelling to begin with. 2. **Prompt Engineering Challenges**: The [prompting guide](https://docs.vapi.ai/prompting-guide) indicates that effective prompt engineering is crucial for creating successful voice AI agents. Sometimes instructions get lost or deprioritized by the AI, especially when there are multiple competing instructions. 3. **Model Behavior**: As noted in some [community discussions](https://discord.com/channels/1211482211119796234/1211483291191083018/threads/1291058129102049300), AI assistants may sometimes fail to adhere to instructions in the prompt, even when they're explicitly stated. ## How to Investigate Using Call IDs To investigate this issue with your call IDs: 1. **Access Call Logs**: Navigate to the Vapi Dashboard, then find the "logs > calls" section in the left-side navigation bar to locate detailed information about those specific calls, as mentioned in [community discussions](https://discord.com/channels/1211482211119796234/1211483291191083018/threads/1339467226381684789). 2. **Check Transcriptions**: Review the transcriptions to see if "Accelnomics" was correctly transcribed from speech to text.
3. **Examine Prompt Application**: Look for evidence of how the assistant processed the instruction about the spelling of "Accelnomics." ## Potential Solutions Based on [information about handling pronunciation issues](https://discord.com/channels/1211482211119796234/1211483291191083018/threads/1332426179285880863), you might try: 1. **Phonetic Spelling**: Try spelling out "Accelnomics" phonetically in the prompt or breaking it into smaller chunks. 2. **SSML Parsing**: Enable SSML parsing and use prompting keywords to help the model output the correct format, as suggested in [community discussions](https://discord.com/channels/1211482211119796234/1211483291191083018/threads/1338644573207007304). 3. **Prompt Refinement**: The [prompting guide](https://docs.vapi.ai/prompting-guide) recommends iterative refinement of prompts through trial and error to achieve more precise responses. Consider strengthening the instruction about "Accelnomics" or placing it in a more prominent position in the prompt. 4. **Model Selection**: Consider using a more recent model which may improve performance, as mentioned in [community discussions](https://discord.com/channels/1211482211119796234/1211483291191083018/threads/1339689865440268328). Would you like me to elaborate on any of these potential solutions or suggest other approaches to resolve this spelling issue? Sources: - [Factors affecting prompt reliability in voice AI assistants](https://docs.vapi.ai/prompting-guide) - [How voice AI assistants handle spelling errors in prompts](https://discord.com/channels/1211482211119796234/1211483291191083018/threads/1338644573207007304) - [How to use call IDs for analysis in Vapi](https://discord.com/channels/1211482211119796234/1211483291191083018/threads/1339467226381684789) - [How Vapi handles the spelling of Accelnomics](https://docs.vapi.ai/customization/multilingual)
- [How voice AI assistants recognize and apply instructions in prompts](https://discord.com/channels/1211482211119796234/1211483291191083018/threads/1291058129102049300)
p
Hey Sahil, You can use Deepgram keywords so that whenever you speak out the company name, the words get transcribed correctly and then eventually passed correctly to the next model in pipeline. Do give it a try and let me know how it goes for you.