Dropping the vector database hands chunk boundaries to Google's whitespace rule

The chunking decision moved in this build. File Search chunks uploaded documents on whitespace with a token budget ^[7]. That rule belongs to Google. The one input still under the builder's control is the whitespace in the markdown he uploads, and that is where an evening went ^[8].

markitdown gave him headings of a sort, and it also gave him this out of a PDF: `Theubiquityoffrustrating,unhelpfulsoftwareinterfaceshasmotivateddecadesofresearch` ^[8]. Maneshwar wrote that the embedding model does not know what `Theubiquityoffrustrating` is, and neither does the retrieval ^[9]. pdftotext fixed the spacing and threw away every heading, so a 300-page book became one undifferentiated scroll ^[10]. pymupdf4llm reads font sizes to decide what is a heading, keeps bold and italics, and handled the spacing on the same PDF ^[11]. On one essay it found 44 real headings where the others found none ^[12]. "A chunk that starts at a heading is a chunk that means something on its own," he wrote ^[13]. The conversion now sits behind `make process-data`, with EPUBs, PDFs and a sync of the blog repos landing in one markdown tree ^[14].

That count measures the converter. Forty-four headings tell you the chunk boundaries fell on section starts; whether the right chunk comes back for a given question is a separate measurement.

The second thing I would check before adopting this is where the store lives. A File Search store sits inside the Google Cloud project behind the API key that created it ^[15]. Create a store with key A, search it with key B from a different project, and the error tells you nothing useful: as far as key B is concerned, the store is simply not there ^[15]. Free-tier per-minute caps are real, and the usual fix is rotating requests across several keys, which a tool-attached call cannot do ^[16]. So the keys got roles: `keys/store.txt` lists keys that each own a complete copy of the corpus in their own store, ingest uploads to all of them, and the search call tries store 0 with key 0, falling through to store 1 with key 1 on a quota or auth failure ^[17].

Indexing is the one step you pay for ^[4]. A fallback chain of N keys pays that one-time embedding bill N times, and every corpus update re-indexes N stores ^[19].

The storage cap has slack at this size. A 1 GB free-tier store against 5.3 MB of markdown leaves room for roughly 190 copies of the whole shelf ^[18]. The per-minute rate is the limit the key chain exists to work around ^[16].

For this to transfer, your corpus has to fit under the cap, change rarely enough that re-ingesting every store stays cheap, sit inside the per-minute quota at your question volume, and survive being cut on whitespace against someone else's token budget. Retrieval is free at query time; the chunks the model pulls in are billed as ordinary context tokens on the call you were already making ^[4].