Retrieval-augmented generation over a 137-book corpus

Chunking long-form argument

Fixed-size chunking destroys books. A 400-token window cuts a chapter's premise away from its conclusion, and retrieval then returns a confident fragment that misrepresents the author. Chunk on structural boundaries — section, argument, ritual sequence — and carry a summary of the parent context into each chunk.

• Split on headings and argument units, not token counts
• Prepend parent-context summaries so fragments stay interpretable
• Store title, edition, chapter, and page with every chunk
• Keep an unmodified original alongside every derived embedding

Citation discipline

An answer without a title and page is not usable for a body of work where precision matters. The system should refuse rather than paraphrase when retrieval scores are weak — an honest 'not covered in these texts' is more valuable to a reader than a fluent blend of three unrelated traditions.

Cross-persona retrieval

This corpus spans two author personas and radically different registers: restored journalism on one side, initiatory practice on the other. Retrieval has to respect that boundary. Blending a newspaper column into an answer about ritual sequence is a correctness failure, not a serendipitous connection.