retrieval-generator

RetrievalGenerator Operator Reference

RetrievalGenerator is an async operator. It reads text from one column, collects only truthy values, calls await self.llm_serving.generate_from_input(llm_inputs, self.system_prompt), and writes the returned list into output_key.

See examples/good.md for a valid usage pattern and examples/bad.md for common failure cases.

1. Import

from dataflow.operators.core_text import RetrievalGenerator
from dataflow.serving import LightRAGServing

2. Constructor

RetrievalGenerator(
    llm_serving=serving,
    system_prompt="You are a helpful agent.",
)

Notes

1. The operator does not initialize the serving backend for you.
2. Any serving object used here must already be ready before run() starts.
3. Default recommendation: use LightRAGServing.

3. Default LightRAGServing Initialization

If you use the default backend, initialize it like this before constructing RetrievalGenerator:

llm_serving = await LightRAGServing.create(
    api_url="https://api.openai.com/v1",
    llm_model_name="gpt-4o",
    embed_model_name="bge-m3:latest",
    embed_binding_host="http://localhost:11434",
    document_list=["knowledge_base.txt"],
)
if llm_serving is None:
    raise RuntimeError("LightRAGServing initialization failed.")

• LightRAGServing.__init__() accepts api_url, key_name_of_api_key, llm_model_name, embed_model_name, embed_binding_host, embedding_dim, max_embed_tokens, and document_list.
• LightRAGServing.create(...) builds self.rag and loads documents.
• DF_API_KEY must exist in the environment, otherwise construction raises ValueError.
• If document loading fails inside create(...), it logs the error and returns None.