pipelines

Pipelines

TowheePipelines

The TowheePipelines prepares insert and search pipeline for the system. It should have methods below to adapt operations in chatbot:

Parameters:

• llm_src: A string to indicate which llm service to use. Supported value:
  • openai
  • dolly
  • ernie
  • dashscope
  • minimax
  • chatglm
  • skychat

Properties:

• search_pipeline: A Towhee pipeline searches relevant information across the project knowledge for the user's query, and then passes both user query and retrieved documents to LLM service to generate the final answer.
• insert_pipeline: A Towhee pipeline firstly loads & splits data from source (URL or file path), and then save documents & corresponding data such as text embeddings in database(s).

By default, it uses Zilliz Cloud or Milvus to store documents with embeddings. If scalar store is enabled, it will use Elastic as default scalar store. You can modify config.py to configure it.

Prompt

• SYSTEM_PROMPT: The system message will be passed to LLM service.

• QUERY_PROMPT: The prompt template will be used to process a user's query.

• PROMPT_OP:

  The function to prepare messages for a LLM operator, which is used in search pipeline. By default, it is operator prompt/template applied with SYSTEM_PROMPT and QUERY_PROMPT.