Group analysis: what each module produces

When you run analysis on a group, outputs fall into four product-facing classes. The same module id can mean different things depending on whether you run on one transcript or a group.

All registered pipeline modules support group runs (supports_group=true) and have an aggregation entry. BERTopic (bertopic) is a registered module whose native stack is optional ([bertopic] / [full] / Docker — see bertopic_optional_module.md); group aggregation refits from pooled source segments (member topic IDs are not joined across fits).

Four output classes

Class

What the user gets

How it is implemented

Group charts (registry-backed)

Aggregate charts under the group run directory, tagged group_aggregate in the manifest

GROUP_CHART_REGISTRY + run_group_aggregate_charts

Group visuals (module-specific path)

Charts that are not produced by GROUP_CHART_REGISTRY but still appear in the group run

wordclouds — charts written inside _aggregate_wordclouds via run_group_wordclouds, tagged group_aggregate and group_visual_special_path

Group data outputs only

CSV/JSON row bundles (and similar) for the group, no group-level chart generator

temporal_dynamics, insight_eligibility, voice_contours, llm_speaker_summary, contextual_emotion, fine_grained_emotion, llm_custom_qa — aggregated in build_registry but omitted from GROUP_CHART_REGISTRY (by design; see registry comment; llm_custom_qa group chart pending v2 loader). Use per-session runs or reports for member charts where applicable.

Group composite / blob only

A JSON blob artifact, not the standard row aggregation + chart pass

summary, llm_summary, narrative_summary, transcript_output — output_type="blob" in the aggregation registry. For llm_summary / llm_speaker_summary, group finalize may additionally run group LLM synthesis (generation-scoped ACTIVE/COMMIT under .group_llm_synthesis/).

Examples by module

Module

Class

stats, sentiment, acts, ner, lexical_diversity, simplified_transcript, transcript_quality, epistemic_markers, politeness, keyphrases, topic_shift, …

Registry-backed group charts (+ row CSV/JSON); transcript_quality pools only within matching ASR provenance cohorts (see below); keyphrases pools noun_chunks by canonical_key (YAKE/KeyBERT deferred) with chart keyphrases.phrases.global

wordclouds

Module-specific group visuals (special path); includes pooled keyphrase_noun_chunks cloud when keyphrase_noun_chunk_pool is passed (never concat-reparse)

temporal_dynamics, insight_eligibility, voice_contours, llm_speaker_summary, contextual_emotion, fine_grained_emotion, llm_custom_qa

Data outputs only (no group chart registry entry); llm_speaker_summary may also feed group LLM synthesis

summary, llm_summary, narrative_summary, transcript_output

Blob-only composite; llm_summary may also feed group LLM synthesis

llm_action_items, insights, semantic_similarity*, voice_mismatch/tension/fingerprint

Rows (+ charts where registered)

prosody (from voice_features / voice_charts_core / prosody_dashboard)

Registry-backed via existing prosody agg

* semantic_similarity, semantic_similarity_advanced, and semantic_similarity share one aggregation entry. B14: v2 motif envelope + group centroid match within a comparable provenance cohort (valid-zero sessions included; TF-IDF incomparable). repetition_rows remain content_rows; additive motif_rows + versioned semantic_similarity_pooled. Chart class: composite session bars + motif prevalence (group_charts_semantic_motifs_contract.md).

keyphrases group aggregation (B16)

  • Pool per-transcript noun_chunk ranks by canonical_key (never reparse concatenated multi-session text).

  • Aggregate: sum occurrence_count, sum member rank_weight then re-normalise, member_session_support, display = mode.

  • Drop phrases below min_member_sessions (default 2).

  • Registry chart: keyphrases.phrases.global (top-N pooled rank_weight); requires keyphrases_pooled on chart_outcome allowlist.

  • YAKE/KeyBERT: explicitly deferred for group pooling this wave.

  • Wordclouds special-path: pooled noun_chunk cloud from explicit keyphrase_noun_chunk_pool into run_group_wordclouds (see pooled_variants.py).

  • Runtime note: ../runtime/keyphrases.md.

  • Pooled chart contract: group_charts_keyphrases_pooled_contract.md.

transcript_quality group aggregation

ASR confidence is not comparable across ASR engines/models/normalisation policies.

  • Members are grouped by provenance.comparable_key.

  • The primary cohort is the largest compatible set; other members are counted as incompatible_member_count and excluded from pooled confidence metrics.

  • Within a cohort (word-weighted, never unweighted mean of session summaries):

    • coverage = sum(scored) / sum(eligible)

    • mean_score = sum(mean_i * scored_i) / sum(scored_i)

    • low_score_ratio = sum(low_score_words) / sum(scored)

  • Group charts emit session bars only for the primary cohort; they never blend incompatible provenance.

See also runtime/transcript_quality.md.

Insights / Overview UI (group subjects)

Pipeline already runs selected modules on each member. The web Insights and Overview blocks load both:

  1. Group rollups (*_rows.json, summary blobs, optional group LLM synthesis)

  2. Per-session member contracts via a session picker (storage_root-aware loader)

See docs/dev/web_blocks.md and docs/dev/group_functionality_audit_2026-07-17.md (E6/E7).