ADR-010: Hierarchical Summarization Pattern¶
- Status: Accepted
- Date: 2026-01-11
- Authors: Podcast Scraper Team
- Related RFCs: RFC-012
- Related PRDs: PRD-005
Context & Problem Statement¶
Most local transformer models (like BART) have a strict context window of 1024 tokens. Podcast transcripts often exceed 10,000 tokens, making direct summarization impossible.
Decision¶
We implement a Hierarchical Summarization Pattern (Map-Reduce):
- Map: Split the transcript into semantic chunks that fit within the model's context window.
- Summarize: Summarize each chunk individually.
- Reduce: Combine the chunk summaries and summarize them again into a final episode abstract.
Rationale¶
- Feasibility: Allows processing transcripts of any length on hardware with limited VRAM.
- Context Preservation: Ensures that details from the beginning, middle, and end of a long podcast are all reflected in the final summary.
Alternatives Considered¶
- Truncation: Summarizing only the first 1024 tokens; rejected as it misses the bulk of the content.
- Long-Context Models (LED): Supported as an alternative, but the hierarchical pattern remains the "safe" default for standard models.
Consequences¶
- Positive: Handles 2+ hour episodes reliably.
- Negative: Slower than single-pass summarization; requires careful management of "summary of summaries" to avoid losing detail.