booksmart-core 0.3.0

Booksmart's book-ingestion pipeline: parsing, structure, extraction, summaries, embeddings — as durable Stage functions consumers drive with their own Runner.

booksmart-core

Booksmart's book-ingestion pipeline as a typed library: parsing, structure detection, profile generation, knowledge extraction, summaries, and embeddings — exposed as durable, synchronous Stage functions that a consumer drives with its own Runner (see ADR 0002).

Core owns the domain: the ORM models, a single dialect-neutral Alembic history (SQLite and Postgres), object storage, the provider abstractions (Anthropic / OpenAI / Gemini, plus deterministic fakes), and the Qdrant vector store. It reads no environment at all — a consumer constructs Settings explicitly (API keys included; since 0.2.0 there is no env-var fallback) and owns engine and session lifecycle.

Consumers:

• booksmart — the local CLI (SQLite, embedded Qdrant).

from booksmart_core.database import create_engine, upgrade_to_head
from booksmart_core.runner import execute_run

upgrade_to_head(url)
run_id = execute_run(session_factory, storage_root, book_id, "full")

Search

booksmart_core.search.search is the read side of the embedding pipeline, and the single seam every consumer's search surface sits on — the CLI's search command calls it against embedded Qdrant; a server consumer would call it with the same arguments against a served instance.

from booksmart_core.search import search

results = search(session, vector_store, embedder, "how do deep modules help?", limit=5)
for hit in results.hits:
    print(hit.score, hit.title)
print(results.embedding_tokens)  # what the query cost; None if the provider withheld it

It embeds the query with the collection's locked model (refusing a mismatch, ADR 0001), ANN-searches Qdrant, and resolves each hit back to its detached ORM row. A search is one unbatched embedding call, so embedding_tokens is the only place a consumer costing search traffic can get that number.