A local-first pipeline for turning a personal book collection into structured, searchable knowledge.
Books are easy to collect and difficult to work with. This build explores a durable path from messy files to catalogued text, summaries, topics, relationships, and grounded questions.
The pipeline proceeds in resumable phases: inventory and metadata recovery, canonical naming, text extraction, summaries, offline topic classification, graph export, embeddings, cross-book relationships, and an assistant surface.
books → catalog → text corpus → summaries/topics
↘ graph → embeddings → grounded assistant| Evidence | What the repository shows |
|---|---|
| Resumable phases | Ten documented phases cover ingestion, enrichment, relationships, and assistant access. |
| Local-first stack | SQLite is the canonical catalog; graph export and vector retrieval are separate, inspectable layers. |
| Retrieval design | Embeddings are paired with lexical/search fallbacks rather than treated as the only path. |
| Operational discipline | Source files remain read-only and the pipeline can be interrupted and resumed. |
Every useful answer depends on knowing which source, extraction, and transformation produced it.
Lexical retrieval and offline stages keep the system useful when a model or vector service is unavailable.
Long-running knowledge pipelines become practical when each stage can be inspected and restarted.
A useful internal library is not automatically a public dataset. Sharing the method is safer than exposing the source material.