Skip to content

Analyzer and Processes

The Analyzer orchestrates one or more processes against a dataset, with simple caching and rich result helpers for downstream analysis and visualization.

Module: pulse.analysis.analyzer, pulse.analysis.processes, pulse.analysis.results

Analyzer

from pulse.analysis.analyzer import Analyzer
from pulse.analysis.processes import ThemeGeneration, SentimentProcess

texts = ["I love pizza", "I hate rain"]
processes = [ThemeGeneration(min_themes=2), SentimentProcess()]
with Analyzer(dataset=texts, processes=processes, cache_dir=".pulse_cache") as az:
    results = az.run()

print(results.theme_generation.to_dataframe())
print(results.sentiment.summary())

Constructor parameters: - dataset: Sequence[str] | pandas.Series – Input data. - processes: Sequence[Process] | None – Ordered list of process instances (see below). - fast: bool | None – Default fast/async flag used by processes when not specified on the process. - cache_dir: str | None – If set and use_cache=True, stores per‑process results in diskcache. - use_cache: bool – Enable/disable caching (default True). - client: CoreClient | None – CoreClient instance to use; defaults to CoreClient(auth=auth). - auth: pulse.auth._BaseOAuth2Auth | None – Auth to pass to an internally constructed CoreClient.

Methods: - run() -> AnalysisResult – Executes processes in order, returning a container exposing each process result as an attribute using the process id / alias. - clear_cache() – Clears on‑disk cache (if enabled). - close() – Closes the underlying client and cache.

Dependency resolution: - Some processes depend on others (e.g., ThemeAllocation depends on ThemeGeneration). The Analyzer automatically inserts missing dependencies with sensible defaults.

Processes (module: pulse.analysis.processes)

All processes implement a common Process protocol: id (str), depends_on (tuple of ids), and run(ctx).

ThemeGeneration

Clusters texts into latent themes.

Parameters: - min_themes: int = 2, max_themes: int = 50 – Bounds for clustering. - context: Any = None – Optional guiding context. - version: str | None = None – Model version pin. - prune: int | None = None – Drop N lowest‑frequency themes. - fast: bool | None = None – Overrides Analyzer’s default fast flag. - await_job_result: bool = True – Return Job when False.

Result wrapper: ThemeGenerationResult (see below).

SentimentProcess

Classifies sentiment for each text.

Parameters: - fast: bool | None = None

Result wrapper: SentimentResult.

ThemeAllocation

Allocates each text to themes by computing similarity between texts and theme labels, then choosing the most similar themes.

Parameters: - themes: list[str] | None = None – Use explicit themes; otherwise use themes from the most recent ThemeGeneration or a specified source. - single_label: bool = True – If true, assign only top theme above threshold. - fast: bool | None = None - threshold: float = 0.5 – Minimum similarity for assignment.

Returns a raw dict in Analyzer ({"themes", "assignments", "similarity"}), which the Analyzer wraps as ThemeAllocationResult.

ThemeExtraction

Extracts elements matching themes from inputs.

Parameters: - themes: list[str] | None = None - version: str | None = None - fast: bool | None = None - dictionary: dict[str, list[str]] | None = None - expand_dictionary: bool | None = None - use_ner: bool | None = None - use_llm: bool | None = None - threshold: float | None = None

Result wrapper: ThemeExtractionResult.

Cluster

Computes a similarity matrix suitable for downstream clustering (performed client‑side).

Parameters: - fast: bool | None = None

Result wrapper: ClusterResult.

Result Helpers (module: pulse.analysis.results)

ThemeGenerationResult

  • themes: list[Theme] – Structured themes.
  • to_dataframe() -> pandas.DataFrame – Columns: shortLabel, label, description, representative_1, representative_2.

SentimentResult

  • sentiments: list[SentimentResultModel] – Sentiment label + confidence per input.
  • to_dataframe() -> pandas.DataFrame – Columns: text, sentiment, confidence.
  • summary() -> pandas.Series – Counts per sentiment label.
  • plot_distribution(**kwargs) – Matplotlib bar chart of the distribution.

ThemeAllocationResult

  • assign_single(threshold: float | None = None) -> pandas.Series – Single best theme per text above threshold (or None).
  • assign_multi(k: int | None = None) -> pandas.DataFrame – Top‑k theme labels per text, ordered by similarity.
  • bar_chart(**kwargs) – Horizontal bar chart of assignment counts.
  • heatmap(**kwargs) – Heatmap of the similarity matrix (uses seaborn if available, falls back to matplotlib).
  • to_dataframe() -> pandas.DataFrame – Long format: text, theme, score.

ClusterResult

  • matrix – NumPy array view of the similarity matrix.
  • kmeans(n_clusters, **kwargs) – Runs scikit‑learn KMeans on the matrix.
  • dbscan(eps=0.4, min_samples=5, **kwargs) – Runs scikit‑learn DBSCAN on a distance transform.
  • plot_scatter(**kwargs) – 2D PCA scatter plot.
  • dendrogram(**kwargs) – Hierarchical clustering dendrogram (SciPy).

ThemeExtractionResult

  • extractions: list[list[list[str]]] – Nested extractions per text per theme.
  • to_dataframe() -> pandas.DataFrame – Long format: text, category, extraction.

Optional Dependencies for Plots/ML

Some helper methods use third‑party libraries: - matplotlib – plotting utilities across results - seaborn – heatmap in ThemeAllocationResult.heatmap() (falls back if missing) - numpy – matrix handling in ClusterResult - scikit‑learn – kmeans, dbscan, and plot_scatter (PCA) - scipy – dendrogram()

Install these as needed, for example:

pip install matplotlib seaborn numpy scikit-learn scipy