SquadAgent¶

The SquadAgent is a question-answering agent that uses a model trained on the Stanford Question Answering Dataset (SQuAD) to extract answers from provided context.

Purpose¶

This agent provides extractive question-answering capabilities within a Rustic AI guild. Given a context passage and a question, it can locate and extract the relevant answer from the context, enabling precise information retrieval.

When to Use¶

Use the SquadAgent when your application needs to:

Extract specific information from longer text passages
Answer factual questions based on provided context
Implement extractive question-answering functionality
Retrieve precise answers rather than generated responses
Process documents and extract key information

Message Types¶

Input Messages¶

QuestionWithContext¶

A request containing a question and its context:

class QuestionWithContext(BaseModel):
    context: str  # The text passage containing the answer
    question: str  # The question to answer

Output Messages¶

SquadAnswer¶

The answer extracted from the context:

class SquadAnswer(BaseModel):
    text: str  # The extracted answer text

ErrorMessage¶

Sent when question answering fails:

class ErrorMessage(BaseModel):
    agent_type: str
    error_type: str  # "InferenceError"
    error_message: str

Behavior¶

The agent receives a QuestionWithContext message with a context passage and a question
It formats the input for the SQuAD model (deepset/roberta-base-squad2)
The request is sent to the Hugging Face Inference API
The model extracts the most relevant answer span from the context
The agent returns a SquadAnswer with the extracted text
If any errors occur, an ErrorMessage is sent

Sample Usage¶

from rustic_ai.core.guild.builders import AgentBuilder
from rustic_ai.huggingface.agents.nlp.squad_agent import SquadAgent

# Create the agent spec
squad_agent_spec = (
    AgentBuilder(SquadAgent)
    .set_id("qa_agent")
    .set_name("Question Answering")
    .set_description("Extracts answers from context using SQuAD")
    .build_spec()
)

# Add to guild
guild_builder.add_agent_spec(squad_agent_spec)

Example Request¶

from rustic_ai.huggingface.agents.nlp.squad_agent import QuestionWithContext

# Create a question-answering request
qa_request = QuestionWithContext(
    context="The Taj Mahal, found in the Indian state of Uttar Pradesh, is one of the Seven Wonders of the Modern World. It was built in the 17th century by Mughal emperor Shah Jahan as a mausoleum for his third wife, Mumtaz Mahal. The Taj Mahal is a fine example of Mughal architecture, which is a blend of Islamic, Persian, Turkish and Indian architectural styles.",
    question="Who was Mumtaz Mahal?"
)

# Send to the agent
client.publish("default_topic", qa_request)

Example Response¶

The agent responds with a SquadAnswer:

SquadAnswer(
    text="his third wife"
)

Technical Details¶

The agent leverages: - Hugging Face's Inference API - deepset/roberta-base-squad2 model, a RoBERTa model fine-tuned on SQuAD 2.0 - Inherits from HuggingfaceInferenceMixin to handle API communication

Notes and Limitations¶

Requires a Hugging Face API key set as the HUGGINGFACE_API_KEY environment variable
Only performs extractive QA (extracts spans from the context, doesn't generate new text)
Performance depends on the relevance of the provided context
Questions must be answerable from the given context
The model works best with factual, specific questions
Context length is limited; very long contexts may need to be split
Performance may vary based on context complexity and question type
The agent operates in REMOTE mode, relying on the Hugging Face API rather than local inference