Idea: raw text snippet windows as profiler context/double-check #36
Description
Idea (to explore)
The profiler currently only sees what the analyzer extracted (structured fragments). If the analyzer missed a subtle mention or miscategorized a role, the profiler never sees it.
Idea: for each character, scan raw source files for all occurrences of their name, extract ±200 chars of surrounding text, and pass those snippets as additional context to the profiler.
When
Likely best as a post-profiling double-check pass rather than upfront — the profiler already has a consolidated profile, the raw snippets are used to validate/enrich rather than flooding it with noisy raw text from the start.
Potential value
- Catches what the structured pipeline lost
- Cheap to compute (regex + slice)
- Could surface missed relations, traits, background details
Open questions
- How to handle overlapping windows (name appears multiple times in close proximity)?
- Noise from stage directions mixed with dialogue
- Same name potentially referring to different characters across eras