Opyt vs building it yourself

The retrieval is a weekend. The ingesters and the calibration are the months.

For a competent engineer this is the alternative that actually holds up, and it deserves a straight answer rather than a sales one. You could build this.

What you would be building is not the retrieval, since hybrid search over SQLite is a weekend. It is the acquisition layer and the accumulated calibration: X pagination that terminates correctly, browser-session capture, cross-platform identity verification through links a person attests to, PDF extraction floors per page, spend previews, fail-safe skip semantics, and ranking floors that were measured rather than guessed.

At a glance

Where building it yourself is better

• It fits exactly your sources and nothing else.
• If X and Substack are not among your sources, the hard part disappears and the build is small.
• No inherited opinions: Opyt insists on person-level ingestion, no synthesis layer and no authoring.
• You learn the domain properly, which is worth something on its own.

Where Opyt is better

• It exists today, it is MIT, and you can read every line before trusting it.
• A corpus you start now is empty. Opyt backfills years of your own history on the first import.
• X pagination terminates on 'this page added no new item', not on the cursor, because X returns a fresh bottom cursor forever, so a cursor-driven loop never ends. A page cap of 100 is a runaway backstop that logs loudly; it is never a silent stop.
• Entity resolution merges two profiles only when one's attested links hit the other's self-declared set, by union-find. It never merges on shared attestations alone, which is the defense against someone linking to a person they are not.
• PDF extraction runs two floors, because they answer different questions: 500 characters total asks whether there is a body at all, and 200 characters per page asks whether the whole document arrived. Total length cannot tell a 20-page scan whose only text layer is a watermark from a one-page note, and a false “complete” is permanent under dedup. The per-page value came from 505 live PDFs where the distribution is bimodal: 15 image-only scans at 29-31 characters per page, then nothing at all until 787.7, then 490 genuine documents up to 8,089. Any value between 100 and 500 classifies all 505 identically.
• Every paid path previews before it spends: confirm=False is the default and writes nothing. Every failure skips rather than writes. A partial vector list propagates so the caller drops that document whole.

Which one

Fork it if it is close. What the months buy you is the ingesters and the calibration; the search itself is the cheap part.

install Opyt free, MIT, five minutes