themefinder 0.6.2__py3-none-any.whl → 0.7.0__py3-none-any.whl

themefinder/__init__.py

@@ -1,18 +1,24 @@
 from .core import (
     find_themes,
     sentiment_analysis,
+    theme_clustering,
     theme_condensation,
     theme_generation,
     theme_mapping,
     theme_refinement,
+    theme_target_alignment,
+    detail_detection,
 )
 
 __all__ = [
     "find_themes",
     "sentiment_analysis",
-    "theme_generation",
+    "theme_clustering",
     "theme_condensation",
-    "theme_refinement",
+    "theme_generation",
     "theme_mapping",
+    "theme_refinement",
+    "theme_target_alignment",
+    "detail_detection",
 ]
 __version__ = "0.1.0"

themefinder/core.py

@@ -3,22 +3,33 @@ from pathlib import Path
 
 import pandas as pd
 from langchain_core.prompts import PromptTemplate
-from langchain_core.runnables import Runnable
-
-from .llm_batch_processor import batch_and_run, load_prompt_from_file
-from .models import SentimentAnalysisOutput, ThemeMappingOutput
-from .themefinder_logging import logger
+from langchain.schema.runnable import RunnableWithFallbacks
+
+from themefinder.llm_batch_processor import batch_and_run, load_prompt_from_file
+from themefinder.models import (
+    SentimentAnalysisResponses,
+    ThemeGenerationResponses,
+    ThemeCondensationResponses,
+    ThemeRefinementResponses,
+    ThemeMappingResponses,
+    DetailDetectionResponses,
+    HierarchicalClusteringResponse,
+    ThemeNode,
+)
+from themefinder.theme_clustering_agent import ThemeClusteringAgent
+from themefinder.themefinder_logging import logger
 
 CONSULTATION_SYSTEM_PROMPT = load_prompt_from_file("consultation_system_prompt")
 
 
 async def find_themes(
     responses_df: pd.DataFrame,
-    llm: Runnable,
+    llm: RunnableWithFallbacks,
     question: str,
     target_n_themes: int | None = None,
     system_prompt: str = CONSULTATION_SYSTEM_PROMPT,
     verbose: bool = True,
+    concurrency: int = 10,
 ) -> dict[str, str | pd.DataFrame]:
     """Process survey responses through a multi-stage theme analysis pipeline.
 
@@ -32,7 +43,7 @@ async def find_themes(
 
     Args:
         responses_df (pd.DataFrame): DataFrame containing survey responses
-        llm (Runnable): Language model instance for text analysis
+        llm (RunnableWithFallbacks): Language model instance for text analysis
         question (str): The survey question
         target_n_themes (int | None, optional): Target number of themes to consolidate to.
             If None, skip theme target alignment step. Defaults to None.
@@ -40,6 +51,7 @@ async def find_themes(
             Defaults to CONSULTATION_SYSTEM_PROMPT.
         verbose (bool): Whether to show information messages during processing.
             Defaults to True.
+        concurrency (int): Number of concurrent API calls to make. Defaults to 10.
 
     Returns:
         dict[str, str | pd.DataFrame]: Dictionary containing results from each pipeline stage:
@@ -56,21 +68,28 @@ async def find_themes(
         llm,
         question=question,
         system_prompt=system_prompt,
+        concurrency=concurrency,
     )
     theme_df, _ = await theme_generation(
         sentiment_df,
         llm,
         question=question,
         system_prompt=system_prompt,
+        concurrency=concurrency,
     )
     condensed_theme_df, _ = await theme_condensation(
-        theme_df, llm, question=question, system_prompt=system_prompt
+        theme_df,
+        llm,
+        question=question,
+        system_prompt=system_prompt,
+        concurrency=concurrency,
     )
     refined_theme_df, _ = await theme_refinement(
         condensed_theme_df,
         llm,
         question=question,
         system_prompt=system_prompt,
+        concurrency=concurrency,
     )
     if target_n_themes is not None:
         refined_theme_df, _ = await theme_target_alignment(
@@ -79,6 +98,7 @@ async def find_themes(
             question=question,
             target_n_themes=target_n_themes,
             system_prompt=system_prompt,
+            concurrency=concurrency,
         )
     mapping_df, mapping_unprocessables = await theme_mapping(
         sentiment_df[["response_id", "response"]],
@@ -86,28 +106,36 @@ async def find_themes(
         question=question,
         refined_themes_df=refined_theme_df,
         system_prompt=system_prompt,
+        concurrency=concurrency,
+    )
+    detailed_df, _ = await detail_detection(
+        responses_df[["response_id", "response"]],
+        llm,
+        question=question,
+        system_prompt=system_prompt,
+        concurrency=concurrency,
     )
 
     logger.info("Finished finding themes")
-    logger.info(
-        "Provide feedback or report bugs: https://forms.gle/85xUSMvxGzSSKQ499 or packages@cabinetoffice.gov.uk"
-    )
+    logger.info("Provide feedback or report bugs: packages@cabinetoffice.gov.uk")
     return {
         "question": question,
         "sentiment": sentiment_df,
         "themes": refined_theme_df,
         "mapping": mapping_df,
+        "detailed_responses": detailed_df,
         "unprocessables": pd.concat([sentiment_unprocessables, mapping_unprocessables]),
     }
 
 
 async def sentiment_analysis(
     responses_df: pd.DataFrame,
-    llm: Runnable,
+    llm: RunnableWithFallbacks,
     question: str,
     batch_size: int = 20,
     prompt_template: str | Path | PromptTemplate = "sentiment_analysis",
     system_prompt: str = CONSULTATION_SYSTEM_PROMPT,
+    concurrency: int = 10,
 ) -> tuple[pd.DataFrame, pd.DataFrame]:
     """Perform sentiment analysis on survey responses using an LLM.
 
@@ -117,7 +145,7 @@ async def sentiment_analysis(
     Args:
         responses_df (pd.DataFrame): DataFrame containing survey responses to analyze.
             Must contain 'response_id' and 'response' columns.
-        llm (Runnable): Language model instance to use for sentiment analysis.
+        llm (RunnableWithFallbacks): Language model instance to use for sentiment analysis.
         question (str): The survey question.
         batch_size (int, optional): Number of responses to process in each batch.
             Defaults to 20.
@@ -126,6 +154,7 @@ async def sentiment_analysis(
             or PromptTemplate instance. Defaults to "sentiment_analysis".
         system_prompt (str): System prompt to guide the LLM's behavior.
             Defaults to CONSULTATION_SYSTEM_PROMPT.
+        concurrency (int): Number of concurrent API calls to make. Defaults to 10.
 
     Returns:
         tuple[pd.DataFrame, pd.DataFrame]:
@@ -134,32 +163,33 @@ async def sentiment_analysis(
                 - The second DataFrame contains the rows that could not be processed by the LLM
 
     Note:
-        The function uses validation_check to ensure responses maintain
+        The function uses integrity_check to ensure responses maintain
         their original order and association after processing.
     """
     logger.info(f"Running sentiment analysis on {len(responses_df)} responses")
-    processed_rows, unprocessable_rows = await batch_and_run(
+    sentiment, unprocessable = await batch_and_run(
         responses_df,
         prompt_template,
-        llm,
+        llm.with_structured_output(SentimentAnalysisResponses),
         batch_size=batch_size,
         question=question,
-        validation_check=True,
-        task_validation_model=SentimentAnalysisOutput,
+        integrity_check=True,
         system_prompt=system_prompt,
+        concurrency=concurrency,
     )
 
-    return processed_rows, unprocessable_rows
+    return sentiment, unprocessable
 
 
 async def theme_generation(
     responses_df: pd.DataFrame,
-    llm: Runnable,
+    llm: RunnableWithFallbacks,
     question: str,
     batch_size: int = 50,
     partition_key: str | None = "position",
     prompt_template: str | Path | PromptTemplate = "theme_generation",
     system_prompt: str = CONSULTATION_SYSTEM_PROMPT,
+    concurrency: int = 10,
 ) -> tuple[pd.DataFrame, pd.DataFrame]:
     """Generate themes from survey responses using an LLM.
 
@@ -168,7 +198,7 @@ async def theme_generation(
     Args:
         responses_df (pd.DataFrame): DataFrame containing survey responses.
             Must include 'response_id' and 'response' columns.
-        llm (Runnable): Language model instance to use for theme generation.
+        llm (RunnableWithFallbacks): Language model instance to use for theme generation.
         question (str): The survey question.
         batch_size (int, optional): Number of responses to process in each batch.
             Defaults to 50.
@@ -181,6 +211,7 @@ async def theme_generation(
             or PromptTemplate instance. Defaults to "theme_generation".
         system_prompt (str): System prompt to guide the LLM's behavior.
             Defaults to CONSULTATION_SYSTEM_PROMPT.
+        concurrency (int): Number of concurrent API calls to make. Defaults to 10.
 
     Returns:
         tuple[pd.DataFrame, pd.DataFrame]:
@@ -193,22 +224,24 @@ async def theme_generation(
     generated_themes, _ = await batch_and_run(
         responses_df,
         prompt_template,
-        llm,
+        llm.with_structured_output(ThemeGenerationResponses),
         batch_size=batch_size,
         partition_key=partition_key,
         question=question,
         system_prompt=system_prompt,
+        concurrency=concurrency,
     )
     return generated_themes, _
 
 
 async def theme_condensation(
     themes_df: pd.DataFrame,
-    llm: Runnable,
+    llm: RunnableWithFallbacks,
     question: str,
     batch_size: int = 75,
     prompt_template: str | Path | PromptTemplate = "theme_condensation",
     system_prompt: str = CONSULTATION_SYSTEM_PROMPT,
+    concurrency: int = 10,
     **kwargs,
 ) -> tuple[pd.DataFrame, pd.DataFrame]:
     """Condense and combine similar themes identified from survey responses.
@@ -219,7 +252,7 @@ async def theme_condensation(
     Args:
         themes_df (pd.DataFrame): DataFrame containing the initial themes identified
             from survey responses.
-        llm (Runnable): Language model instance to use for theme condensation.
+        llm (RunnableWithFallbacks): Language model instance to use for theme condensation.
         question (str): The survey question.
         batch_size (int, optional): Number of themes to process in each batch.
             Defaults to 100.
@@ -228,6 +261,7 @@ async def theme_condensation(
             or PromptTemplate instance. Defaults to "theme_condensation".
         system_prompt (str): System prompt to guide the LLM's behavior.
             Defaults to CONSULTATION_SYSTEM_PROMPT.
+        concurrency (int): Number of concurrent API calls to make. Defaults to 10.
 
     Returns:
         tuple[pd.DataFrame, pd.DataFrame]:
@@ -247,10 +281,11 @@ async def theme_condensation(
         themes_df, _ = await batch_and_run(
             themes_df,
             prompt_template,
-            llm,
+            llm.with_structured_output(ThemeCondensationResponses),
             batch_size=batch_size,
             question=question,
             system_prompt=system_prompt,
+            concurrency=concurrency,
             **kwargs,
         )
         themes_df = themes_df.sample(frac=1).reset_index(drop=True)
@@ -263,10 +298,11 @@ async def theme_condensation(
     themes_df, _ = await batch_and_run(
         themes_df,
         prompt_template,
-        llm,
+        llm.with_structured_output(ThemeCondensationResponses),
         batch_size=batch_size,
         question=question,
         system_prompt=system_prompt,
+        concurrency=concurrency,
         **kwargs,
     )
 
@@ -274,13 +310,95 @@ async def theme_condensation(
     return themes_df, _
 
 
+def theme_clustering(
+    themes_df: pd.DataFrame,
+    llm: RunnableWithFallbacks,
+    max_iterations: int = 5,
+    target_themes: int = 10,
+    significance_percentage: float = 10.0,
+    return_all_themes: bool = False,
+) -> tuple[pd.DataFrame, pd.DataFrame]:
+    """Perform hierarchical clustering of themes using an agentic approach.
+
+    This function takes a DataFrame of themes and uses the ThemeClusteringAgent
+    to iteratively merge similar themes into a hierarchical structure, then
+    selects the most significant themes based on a threshold.
+
+    Args:
+        themes_df (pd.DataFrame): DataFrame containing themes with columns:
+            - topic_id: Unique identifier for each theme
+            - topic_label: Short descriptive label for the theme
+            - topic_description: Detailed description of the theme
+            - source_topic_count: Number of source responses for this theme
+        llm (RunnableWithFallbacks): Language model instance configured with
+            structured output for HierarchicalClusteringResponse
+        max_iterations (int, optional): Maximum number of clustering iterations.
+            Defaults to 5.
+        target_themes (int, optional): Target number of themes to cluster down to.
+            Defaults to 10.
+        significance_percentage (float, optional): Percentage threshold for
+            selecting significant themes. Defaults to 10.0.
+        return_all_themes (bool, optional): If True, returns all clustered themes.
+            If False, returns only significant themes. Defaults to False.
+
+    Returns:
+        tuple[pd.DataFrame, pd.DataFrame]:
+            A tuple containing:
+                - DataFrame of clustered themes (all or significant based on return_all_themes)
+                - Empty DataFrame (for consistency with other functions)
+    """
+    logger.info(f"Starting hierarchical clustering of {len(themes_df)} themes")
+
+    # Convert DataFrame to ThemeNode objects
+    initial_themes = [
+        ThemeNode(
+            topic_id=row["topic_id"],
+            topic_label=row["topic_label"],
+            topic_description=row["topic_description"],
+            source_topic_count=row["source_topic_count"],
+        )
+        for _, row in themes_df.iterrows()
+    ]
+
+    # Initialize clustering agent with structured output LLM
+    agent = ThemeClusteringAgent(
+        llm.with_structured_output(HierarchicalClusteringResponse), initial_themes
+    )
+
+    # Perform clustering
+    logger.info(
+        f"Clustering themes with max_iterations={max_iterations}, target_themes={target_themes}"
+    )
+    all_themes_df = agent.cluster_themes(
+        max_iterations=max_iterations, target_themes=target_themes
+    )
+
+    # Return appropriate themes based on parameter
+    if return_all_themes:
+        logger.info(
+            f"Clustering complete: returning all {len(all_themes_df)} clustered themes"
+        )
+        return all_themes_df, pd.DataFrame()
+    else:
+        # Select significant themes
+        logger.info(
+            f"Selecting themes with significance_percentage={significance_percentage}%"
+        )
+        selected_themes_df = agent.select_themes(significance_percentage)
+        logger.info(
+            f"Clustering complete: returning {len(selected_themes_df)} significant themes"
+        )
+        return selected_themes_df, pd.DataFrame()
+
+
 async def theme_refinement(
     condensed_themes_df: pd.DataFrame,
-    llm: Runnable,
+    llm: RunnableWithFallbacks,
     question: str,
     batch_size: int = 10000,
     prompt_template: str | Path | PromptTemplate = "theme_refinement",
     system_prompt: str = CONSULTATION_SYSTEM_PROMPT,
+    concurrency: int = 10,
 ) -> tuple[pd.DataFrame, pd.DataFrame]:
     """Refine and standardize condensed themes using an LLM.
 
@@ -292,7 +410,7 @@ async def theme_refinement(
     Args:
         condensed_themes (pd.DataFrame): DataFrame containing the condensed themes
             from the previous pipeline stage.
-        llm (Runnable): Language model instance to use for theme refinement.
+        llm (RunnableWithFallbacks): Language model instance to use for theme refinement.
         question (str): The survey question.
         batch_size (int, optional): Number of themes to process in each batch.
             Defaults to 10000.
@@ -301,6 +419,7 @@ async def theme_refinement(
             or PromptTemplate instance. Defaults to "theme_refinement".
         system_prompt (str): System prompt to guide the LLM's behavior.
             Defaults to CONSULTATION_SYSTEM_PROMPT.
+        concurrency (int): Number of concurrent API calls to make. Defaults to 10.
 
     Returns:
         tuple[pd.DataFrame, pd.DataFrame]:
@@ -319,22 +438,24 @@ async def theme_refinement(
     refined_themes, _ = await batch_and_run(
         condensed_themes_df,
         prompt_template,
-        llm,
+        llm.with_structured_output(ThemeRefinementResponses),
         batch_size=batch_size,
         question=question,
         system_prompt=system_prompt,
+        concurrency=concurrency,
     )
     return refined_themes, _
 
 
 async def theme_target_alignment(
     refined_themes_df: pd.DataFrame,
-    llm: Runnable,
+    llm: RunnableWithFallbacks,
     question: str,
     target_n_themes: int = 10,
     batch_size: int = 10000,
     prompt_template: str | Path | PromptTemplate = "theme_target_alignment",
     system_prompt: str = CONSULTATION_SYSTEM_PROMPT,
+    concurrency: int = 10,
 ) -> tuple[pd.DataFrame, pd.DataFrame]:
     """Align themes to target number using an LLM.
 
@@ -346,7 +467,7 @@ async def theme_target_alignment(
     Args:
         refined_themes_df (pd.DataFrame): DataFrame containing the refined themes
             from the previous pipeline stage.
-        llm (Runnable): Language model instance to use for theme alignment.
+        llm (RunnableWithFallbacks): Language model instance to use for theme alignment.
         question (str): The survey question.
         target_n_themes (int, optional): Target number of themes to consolidate to.
             Defaults to 10.
@@ -357,6 +478,7 @@ async def theme_target_alignment(
             or PromptTemplate instance. Defaults to "theme_target_alignment".
         system_prompt (str): System prompt to guide the LLM's behavior.
             Defaults to CONSULTATION_SYSTEM_PROMPT.
+        concurrency (int): Number of concurrent API calls to make. Defaults to 10.
 
     Returns:
         tuple[pd.DataFrame, pd.DataFrame]:
@@ -376,23 +498,25 @@ async def theme_target_alignment(
     aligned_themes, _ = await batch_and_run(
         refined_themes_df,
         prompt_template,
-        llm,
+        llm.with_structured_output(ThemeRefinementResponses),
         batch_size=batch_size,
         question=question,
         system_prompt=system_prompt,
         target_n_themes=target_n_themes,
+        concurrency=concurrency,
     )
     return aligned_themes, _
 
 
 async def theme_mapping(
     responses_df: pd.DataFrame,
-    llm: Runnable,
+    llm: RunnableWithFallbacks,
     question: str,
     refined_themes_df: pd.DataFrame,
     batch_size: int = 20,
     prompt_template: str | Path | PromptTemplate = "theme_mapping",
     system_prompt: str = CONSULTATION_SYSTEM_PROMPT,
+    concurrency: int = 10,
 ) -> tuple[pd.DataFrame, pd.DataFrame]:
     """Map survey responses to refined themes using an LLM.
 
@@ -402,7 +526,7 @@ async def theme_mapping(
     Args:
         responses_df (pd.DataFrame): DataFrame containing survey responses.
             Must include 'response_id' and 'response' columns.
-        llm (Runnable): Language model instance to use for theme mapping.
+        llm (RunnableWithFallbacks): Language model instance to use for theme mapping.
         question (str): The survey question.
         refined_themes_df (pd.DataFrame): Single-row DataFrame where each column
             represents a theme (from theme_refinement stage).
@@ -413,6 +537,7 @@ async def theme_mapping(
             or PromptTemplate instance. Defaults to "theme_mapping".
         system_prompt (str): System prompt to guide the LLM's behavior.
             Defaults to CONSULTATION_SYSTEM_PROMPT.
+        concurrency (int): Number of concurrent API calls to make. Defaults to 10.
 
     Returns:
         tuple[pd.DataFrame, pd.DataFrame]:
@@ -432,17 +557,70 @@ async def theme_mapping(
         )
         return transposed_df
 
-    mapping, _ = await batch_and_run(
+    mapping, unprocessable = await batch_and_run(
         responses_df,
         prompt_template,
-        llm,
+        llm.with_structured_output(ThemeMappingResponses),
         batch_size=batch_size,
         question=question,
         refined_themes=transpose_refined_themes(refined_themes_df).to_dict(
             orient="records"
         ),
-        validation_check=True,
-        task_validation_model=ThemeMappingOutput,
+        integrity_check=True,
+        system_prompt=system_prompt,
+        concurrency=concurrency,
+    )
+    return mapping, unprocessable
+
+
+async def detail_detection(
+    responses_df: pd.DataFrame,
+    llm: RunnableWithFallbacks,
+    question: str,
+    batch_size: int = 20,
+    prompt_template: str | Path | PromptTemplate = "detail_detection",
+    system_prompt: str = CONSULTATION_SYSTEM_PROMPT,
+    concurrency: int = 10,
+) -> tuple[pd.DataFrame, pd.DataFrame]:
+    """Identify responses that provide high-value detailed evidence.
+
+    This function processes survey responses in batches to analyze their level of detail
+    and evidence using a language model. It identifies responses that contain specific
+    examples, data, or detailed reasoning that provide strong supporting evidence.
+
+    Args:
+        responses_df (pd.DataFrame): DataFrame containing survey responses to analyze.
+            Must contain 'response_id' and 'response' columns.
+        llm (RunnableWithFallbacks): Language model instance to use for detail detection.
+        question (str): The survey question.
+        batch_size (int, optional): Number of responses to process in each batch.
+            Defaults to 20.
+        prompt_template (str | Path | PromptTemplate, optional): Template for structuring
+            the prompt to the LLM. Can be a string identifier, path to template file,
+            or PromptTemplate instance. Defaults to "detail_detection".
+        system_prompt (str): System prompt to guide the LLM's behavior.
+            Defaults to CONSULTATION_SYSTEM_PROMPT.
+        concurrency (int): Number of concurrent API calls to make. Defaults to 10.
+
+    Returns:
+        tuple[pd.DataFrame, pd.DataFrame]:
+            A tuple containing two DataFrames:
+                - The first DataFrame contains the rows that were successfully processed by the LLM
+                - The second DataFrame contains the rows that could not be processed by the LLM
+
+    Note:
+        The function uses response_id_integrity_check to ensure responses maintain
+        their original order and association after processing.
+    """
+    logger.info(f"Running detail detection on {len(responses_df)} responses")
+    detailed, _ = await batch_and_run(
+        responses_df,
+        prompt_template,
+        llm.with_structured_output(DetailDetectionResponses),
+        batch_size=batch_size,
+        question=question,
+        integrity_check=True,
         system_prompt=system_prompt,
+        concurrency=concurrency,
     )
-    return mapping, _
+    return detailed, _