LLM de CommPath frente a LLM de vista previa

Note: To generate your extractions, there are two different available LLMs to choose from:

1. CommPath LLM
2. Preview LLM

Below is an outline of some of the considerations when deciding on an LLM to use. If your use case requires extracting more than 30 fields per message, we currently recommend using the Preview LLM.

CommPath LLM

• Leverages UiPath’s proprietary LLM, fine-tuned for Communications data.
• Currently limited to extracting approx. 30 fields per message​.
• Less latency than the Preview LLM​.
• Can be fine-tuned based on your data.
• Improving performance for CommPath, both in terms of the number of fields which can be extracted and the inference speed for the model is a high priority for 2024.
• Higher occurrence confidences (check the Automating with GenEx section for more details) compared to the Preview LLM.​

Preview LLM​

• Leverages Azure OpenAI’s GPT model as the underlying LLM.
• UiPath cannot guarantee uptime, as this is entirely dependent on the Azure OpenAI endpoints. If the endpoints are down or overloaded, UiPath cannot guarantee availability​.
• You can extract more than 30 fields per message.
• Higher amount of latency compared to CommPath LLM​.
• Limited to in-context learning​.
  Note: When using in-context learning, the platform can only learn from what you prompt it with. Communications Mining can automatically refine the prompt to an extent, but the model doesn't learn from any user-led validation.

Use the settings illustrated below, to select which LLM you want to use for the Generative Extraction.

CommPath LLM is enabled by default. To enable the Preview LLM, the toggles from the following image are required.

If the Use preview Generative Extraction model toggle is turned off, it means that you are using the CommPath LLM.

Having the Use generative AI features and Use preview Generative Extraction model toggles turned on means the platform uses the UiPath Azure OpenAI endpoint in the extraction process.

Recommended approach

1. Start training your extractions with the CommPath LLM.
2. If the extractions extract correctly, continue to train the extractions using the CommPath LLM. If not, due to high number of fields or large tables in each message, switch to the Preview LLM. ​

   To determine if your extractions predict correctly, check the validation statistics in the Generative Extraction tab, on the Validation page. If the precision and recall of the extractions are appropriate for your use case, continue to use the CommPath LLM.

   If any data points don't extract as expected with the CommPath LLM:

   1. Pin the current model version by going to models and select pin on the most recent model version.
   2. Reach out to your UiPath Representative, making note of the model version where the extractions were not performing well on. Your UiPath Representative will work directly with the Communications Mining product team to investigate and implement improvements.​
3. If you use the Preview LLM, continue to train your model the same way you trained the CommPath LLM. Go through it, and provide correct examples for each of your extractions.