Readers¶

Generic Reader¶

class pyterrier_rag.readers.Reader(backend, prompt='Use the context information to answer the Question: \\n Context: {{ qcontext }} \\n Question: {{ query }} \\n Answer:', output_field='qanswer')[source]¶

Bases: Transformer

Transformer that generates answers from context and queries using an LLM backend.

Combines a PromptTransformer with a Backend to produce text or logprobs, then applies answer extraction to return final responses.

Parameters:

backend (Backend or str) – A Backend instance or model identifier string.
prompt (PromptTransformer or str) – Prompt template or raw instruction.
output_field (str) – Field name in the output DataFrame for answers.

Raises:

ValueError – If the prompt expects logprobs but the backend does not support logprobs.

Example using a local LLM:

from pyterrier_rag.backend import Seq2SeqLMBackend
from pyterrier_rag.prompt import Concatenator
from pyterrier_rag.readers import Reader

flant5 = Reader(Seq2SeqLMBackend('google/flan-t5-base'))
bm25_flant5 = bm25_ret % 10 >> Concatenator() >> flant5
bm25_flant5.search("What is the capital of France?")

Example using a remote LLM:

from pyterrier_rag.backend import OpenAIBackend
from pyterrier_rag.prompt import Concatenator
from pyterrier_rag.readers import Reader

llamma = Reader(OpenAIBackend("llama-3-8b-instruct", api_key="your_api_key", base_url="your_base_url"))
bm25_llamma = bm25_ret % 10 >> Concatenator() >> llamma
bm25_llamma.search("What is the capital of Italy?")

transform(inp)[source]¶

Return type:: DataFrame
Parameters:: inp (DataFrame)

Specific Readers¶

class pyterrier_rag.readers.T5FiD(model_name_or_path, tokenizer_name_or_path=None, batch_size=4, text_field='text', text_max_length=256, num_context='auto', max_new_tokens=32, generation_config=None, verbose=False, device=None, **kwargs)[source]¶

T5 FiD Reader for PyTerrier-RAG

Citation

Izacard and Grave. Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering. EACL 2021. [link]

@inproceedings{DBLP:conf/eacl/IzacardG21,
  author       = {Gautier Izacard and
                  Edouard Grave},
  editor       = {Paola Merlo and
                  Jörg Tiedemann and
                  Reut Tsarfaty},
  title        = {Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering},
  booktitle    = {Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, EACL 2021, Online, April 19 - 23, 2021},
  pages        = {874--880},
  publisher    = {Association for Computational Linguistics},
  year         = {2021},
  isbn         = {978-1-954085-02-2},
  url          = {https://doi.org/10.18653/v1/2021.eacl-main.74},
  doi          = {10.18653/v1/2021.eacl-main.74},
  timestamp    = {2022-01-20},
  bibsource    = {dblp computer science bibliography, https://dblp.org}
}

Parameters:

model_name_or_path (str)
tokenizer_name_or_path (str)
batch_size (int)
text_field (str)
text_max_length (int)
num_context (int | str)
max_new_tokens (int)
generation_config (GenerationConfig)
verbose (bool)
device (str | device)

class pyterrier_rag.readers.BARTFiD(model_name_or_path, tokenizer_name_or_path=None, batch_size=4, text_field='text', text_max_length=256, num_context='auto', max_new_tokens=32, generation_config=None, verbose=False, device=None, **kwargs)[source]¶

BART FiD Reader for PyTerrier-RAG

Citation

Izacard and Grave. Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering. EACL 2021. [link]

@inproceedings{DBLP:conf/eacl/IzacardG21,
  author       = {Gautier Izacard and
                  Edouard Grave},
  editor       = {Paola Merlo and
                  Jörg Tiedemann and
                  Reut Tsarfaty},
  title        = {Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering},
  booktitle    = {Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, EACL 2021, Online, April 19 - 23, 2021},
  pages        = {874--880},
  publisher    = {Association for Computational Linguistics},
  year         = {2021},
  isbn         = {978-1-954085-02-2},
  url          = {https://doi.org/10.18653/v1/2021.eacl-main.74},
  doi          = {10.18653/v1/2021.eacl-main.74},
  timestamp    = {2022-01-20},
  bibsource    = {dblp computer science bibliography, https://dblp.org}
}

Parameters:

model_name_or_path (str)
tokenizer_name_or_path (str)
batch_size (int)
text_field (str)
text_max_length (int)
num_context (int | str)
max_new_tokens (int)
generation_config (GenerationConfig)
verbose (bool)
device (str | device)