Narrative Generation After Data Analysis: Exploring Alternatives

[mraguth]

I like to write narratives in human language after analyzing structured data using a pandas query pipeline. Currently, I am using response_synthesis_prompt for this task. Are there similar research agents in other packages? Specifically, can a research agent review the results and generate a narrative? Is this possible with LlamaIndex?

If, yes, can you give me a sample code please?

[dosubot[bot]]

Yes, it is possible to use LlamaIndex to review the results and generate a narrative. Below is a sample code using PandasQueryEngine from LlamaIndex to achieve this:

import pandas as pd
from llama_index.experimental.query_engine.pandas import PandasQueryEngine
from llama_index.core.response_synthesizers.factory import get_response_synthesizer
from llama_index.core.response_synthesizers.type import ResponseMode

# Sample DataFrame
df = pd.DataFrame(
    {
        "city": ["Toronto", "Tokyo", "Berlin"],
        "population": [2930000, 13960000, 3645000]
    }
)

# Initialize the PandasQueryEngine
query_engine = PandasQueryEngine(df=df, verbose=True)

# Get a response synthesizer for narrative generation
response_synthesizer = get_response_synthesizer(response_mode=ResponseMode.GENERATION)

# Query the DataFrame and generate a narrative
response = query_engine.query("What is the population of Tokyo?", response_synthesizer=response_synthesizer)
print(response.response)

In this example, the get_response_synthesizer function is used to create a response synthesizer with the ResponseMode.GENERATION mode to enable the generation of a narrative response. The query method is then used to ask a question about the DataFrame, and the response will include a synthesized narrative based on the query results ^[1][2].

To continue talking to Dosu, mention @dosu.

[mraguth]

@dosu Thank you. What is the difference between 'response_synthesis_prompt' and 'response_synthesizer'? I used prompt to write about my dataset and attributes.

[dosubot[bot]]

The response_synthesis_prompt and response_synthesizer in LlamaIndex serve different purposes:

1. response_synthesis_prompt:

   • Refers to the templates or prompts used to guide the synthesis process.
   • These prompts can be customized or default ones like DEFAULT_TEXT_QA_PROMPT_SEL, DEFAULT_REFINE_PROMPT_SEL, and DEFAULT_TREE_SUMMARIZE_PROMPT_SEL.
   • They help structure the questions and refine the responses during the synthesis process.

2. response_synthesizer:

   • A comprehensive component that uses the prompts (like response_synthesis_prompt) along with other configurations to generate the final narrative.
   • Instantiated using the get_response_synthesizer function, which takes various parameters including the LLM, prompt templates, response mode, and other settings.
   • Different response modes (e.g., COMPACT, REFINE, TREE_SUMMARIZE, SIMPLE_SUMMARIZE, etc.) determine the specific synthesizer class to be used (e.g., CompactAndRefine, Refine, TreeSummarize, SimpleSummarize).

In summary, response_synthesis_prompt refers to the templates guiding the synthesis, while response_synthesizer is the component that uses these prompts and other settings to generate the final narrative ^[1][2].

To continue talking to Dosu, mention @dosu.

[mraguth]

@dosu. Is that ok to use both in my query? or any downside?

[dosubot[bot]]

Yes, it is okay to use both response_synthesis_prompt and response_synthesizer in your query. There are no explicit downsides mentioned in the provided documentation. The response_synthesizer is a flexible component that can be configured with various modes and custom prompts, including additional variables during query-time.

For example, you can customize the prompts used in the response synthesizer and add additional variables:

from llama_index.core import PromptTemplate
from llama_index.core.response_synthesizers import TreeSummarize

# NOTE: we add an extra tone_name variable here
qa_prompt_tmpl = (
    "Context information is below.\n"
    "---------------------\n"
    "{context_str}\n"
    "---------------------\n"
    "Given the context information and not prior knowledge, "
    "answer the query.\n"
    "Please also write the answer in the tone of {tone_name}.\n"
    "Query: {query_str}\n"
    "Answer: "
)
qa_prompt = PromptTemplate(qa_prompt_tmpl)

# initialize response synthesizer
summarizer = TreeSummarize(verbose=True, summary_template=qa_prompt)

# get response
response = summarizer.get_response(
    "who is Paul Graham?", [text], tone_name="a Shakespeare play"
)

This example shows how you can use custom prompt templates with additional variables in the response synthesizer.

The response_synthesizer can be configured with different modes such as REFINE, COMPACT, TREE_SUMMARIZE, SIMPLE_SUMMARIZE, GENERATION, ACCUMULATE, COMPACT_ACCUMULATE, and NO_TEXT. Each mode has its own specific use case and configuration, allowing for a high degree of flexibility depending on your needs ^[1][2].

To continue talking to Dosu, mention @dosu.