autorag.nodes.generator package

Submodules

autorag.nodes.generator.base module

class autorag.nodes.generator.base.BaseGenerator(project_dir: str, llm: str, *args, **kwargs)

Bases: BaseModule

abstract async astream(prompt: str, **kwargs)

cast_to_run(previous_result: DataFrame, *args, **kwargs)

This function is for cast function (a.k.a decorator) only for pure function in the whole node.

abstract stream(prompt: str, **kwargs)

structured_output(prompts: List[str], output_cls)

autorag.nodes.generator.base.generator_node(func)

autorag.nodes.generator.llama_index_llm module

class autorag.nodes.generator.llama_index_llm.LlamaIndexLLM(project_dir: str, llm: str, batch: int = 16, *args, **kwargs)

Bases: BaseGenerator

async astream(prompt: str | List[dict], **kwargs)

get_default_log_probs(tokenized_ids: List[List[int]]) → List[List[float]]

get_default_tokenized_ids(generated_texts: List[str]) → List[List[int]]

pure(previous_result: DataFrame, *args, **kwargs)

stream(prompt: str | List[dict], **kwargs)

autorag.nodes.generator.openai_llm module

class autorag.nodes.generator.openai_llm.OpenAILLM(project_dir, llm: str, batch: int = 16, *args, **kwargs)

Bases: BaseGenerator

async astream(prompt: str | List[Dict], **kwargs)

async get_result(prompt: str | List[dict], **kwargs)

async get_result_gpt_5(prompt: str | List[dict], **kwargs)

async get_result_reasoning(prompt: str | List[dict], **kwargs)

async get_structured_result(prompt: str | List[Dict], output_cls, **kwargs)

pure(previous_result: DataFrame, *args, **kwargs)

stream(prompt: str | List[Dict], **kwargs)

structured_output(prompts: List[str], output_cls, **kwargs)

autorag.nodes.generator.openai_llm.get_max_token_size(llm: str) → int | None

Resolve the context window size for an OpenAI model name.

Exact matches in MAX_TOKEN_DICT win first. Dated snapshots (e.g. gpt-5.6-sol-2026-07-09) fall back to their base model entry, and unknown gpt-5 family variants fall back to the family context window so that newly released snapshots keep working without a code change.

autorag.nodes.generator.openai_llm.parse_prompt(prompt: str | List[Dict]) → List[Dict]

autorag.nodes.generator.openai_llm.tiktoken_messages_to_string(messages: List[Dict[str, str]]) → str

Convert chat messages to string format for accurate token counting

autorag.nodes.generator.openai_llm.truncate_by_token(prompt: str | List[Dict], tokenizer: Encoding, max_token_size: int)

autorag.nodes.generator.run module

autorag.nodes.generator.run.evaluate_generator_node(result_df: DataFrame, metric_inputs: List[MetricInput], metrics: List[str] | List[Dict])

autorag.nodes.generator.run.run_generator_node(modules: List, module_params: List[Dict], previous_result: DataFrame, node_line_dir: str, strategies: Dict) → DataFrame

Run evaluation and select the best module among generator node results. And save the results and summary to generator node directory.

Parameters:

• modules – Generator modules to run.

• module_params – Generator module parameters. Including node parameters, which is used for every module in this node.

• previous_result – Previous result dataframe. Could be prompt maker node’s result.

• node_line_dir – This node line’s directory.

• strategies – Strategies for generator node.

Returns:

The best result dataframe. It contains previous result columns and generator node’s result columns.

autorag.nodes.generator.vllm module

class autorag.nodes.generator.vllm.Vllm(project_dir: str, llm: str, **kwargs)

Bases: BaseGenerator

async astream(prompt: str, **kwargs)

pure(previous_result: DataFrame, *args, **kwargs)

stream(prompt: str, **kwargs)

Module contents