💬 Annotation Reply
What is Annotated Reply
Annotated reply refers to the ability to provide customizable high-quality question and answer capabilities for applications through manually editing Q&A pairs.
Applicable Scenarios of Annotated Reply
- Annotated reply is suitable for customizing answers in specific domains: For example, in customer service Q&A scenarios, for specific questions, if you want to ensure that the system answers with a “standard answer”, or for some questions “do not answer”, you can use the annotated reply ability to customize.
- Rapid tuning of product DEMO: In the POC stage of the product, customized answers can be achieved through annotated replies, to improve the generation expectations of Q&A results and improve the experience.
The annotated reply is equivalent to providing another set of retrieval enhancement systems, bypassing the generation link of the large model, and avoiding the generation illusion problem of RAG.
Practice of Annotated Reply
Before using the annotated reply capability, you need to make the following preparations:
- Prepare Q&A pair data: Provide the question and its standard answer
- Vectorization of Q&A pair: Vectorize the above Q&A pair, for subsequent retrieval
When providing Q&A pair data, you can refer to the following table. In addition to the QA pair, an additional “similar question” column is added to improve the retrieval effect:
Question | Answer | Similar Question 1 | Similar Question 2 |
---|---|---|---|
What is the capital of Zhejiang Province? | Hangzhou | Where is the administrative center of Zhejiang Province? | Where is the provincial government of Zhejiang Province located? |
… | … | … | … |
After completing the above preparations, the following steps can be executed when the user asks questions later:
- Question Retrieval: Vectorize the user’s question and query whether there are similar annotated questions
- Direct Answer: If similar questions exist, directly return the answer corresponding to the question in the annotation, and do not pass it to RAG or LLM
- RAG or LLM Answer: If there are no similar questions, pass the question to RAG or LLM
Last updated: