Here's what we've recently added:

Hallucinations Detection

A new node called HallucinationDetector detects hallucinations in your retrieval augmented generative question answering systems. It identifies the parts of the answer based on your documents, the parts that aren't, and the parts that contradict what's in your documents. This way, you know if the answer is reliable.

Check also our tutorial on how to create a generative QA app that detects hallucinations: Tutorial: Building a Robust RAG System.

Easy Experimenting with Prompts

Say hello to Prompt Explorer your sandbox environment for prompt testing. Compare up to three pipelines and see how your prompt impacts the answers.

Don't know where to start? We've added a library of prompts for various NLP tasks that you can use and modify as you like. You can also save your own prompts for future use.

PromptNode Changes

We've simplified PromptTemplate used by PromptNode. No more hassles with the name and prompt_text parameters. Now it's just prompt and an optional output_parser. Here's how you can now create your own prompt:

components:
  - name: qa_template # This section configures the new prompt template
    type: PromptTemplate
    params:
       output_parser:
         type: AnswerParser
       prompt: "You are a financial analyst. You are provided with the following\
        \ context: \"\"\" {contexts} \"\"\" Please answer the following question truthfully\
        \ based on the context above: {questions} Answer:  \n"
  - name: PromptNode
    type: PromptNode
    params:
        default_prompt_template: qa_template

What changed:

• The prompt_text parameter was replaced with the prompt parameter.
• The name parameter was deleted.

Note that your deployed pipelines will continue running as before. But if you undeploy a pipeline, you'll need to update the Prompt Template to the new format before redeploying.

- name: qa_template # This section configures the new prompt template
  type: PromptTemplate
  params:
     name: qa
     output_parser:
       type: AnswerParser
     prompt_text: "You are a financial analyst. You are provided with the following\
        \ context: \"\"\" {contexts} \"\"\" Please answer the following question truthfully\
        \ based on the context above: {questions} Answer:  \n"

- name: qa_template # This section configures the new prompt template
  type: PromptTemplate
  params:
     output_parser:
       type: AnswerParser
     prompt: "You are a financial analyst. You are provided with the following\
        \ context: \"\"\" {contexts} \"\"\" Please answer the following question truthfully\
        \ based on the context above: {questions} Answer:  \n"

For details, see Prompt Templates.

Models Hosted on SageMaker

You can now use open source large language models hosted on Amazon SageMaker. This way, you can use much larger models and they run in your own AWS account. Contact us: we'll deploy the model in SageMaker and come back with the model name you should pass in the model_name_or_path parameter of PromptNode.