Saytu Blog https://saytu.ai/blog Notes on anime and otaku ecommerce from the team building Saytu. en Five near-matches kept the exact product away from the model https://saytu.ai/blog/five-near-matches-kept-the-exact-product-away-from-the-model https://saytu.ai/blog/five-near-matches-kept-the-exact-product-away-from-the-model Wed, 23 Sep 2026 10:35:11 GMT Five near-matches hid the exact Shopify product from the model. We changed product search to try precise intent first, then rank 25 broad candidates before cutting the list. The AI started talking again. The human request stayed open. https://saytu.ai/blog/the-ai-started-talking-again-the-human-request-stayed-open https://saytu.ai/blog/the-ai-started-talking-again-the-human-request-stayed-open Wed, 23 Sep 2026 07:38:36 GMT Saytu used to leave abandoned handoffs silent forever. We changed the flow so AI can resume after staff silence while the human request stays open. Twenty chunks did not mean we had the whole document https://saytu.ai/blog/twenty-chunks-did-not-mean-we-had-the-whole-document https://saytu.ai/blog/twenty-chunks-did-not-mean-we-had-the-whole-document Wed, 23 Sep 2026 02:55:05 GMT Saytu could not tell whether a source at the 20-chunk cap was complete or truncated. We now record the pre-cap count so coverage is measured instead of guessed. Semantic retrieval was off in production, and the fallback hid it https://saytu.ai/blog/semantic-retrieval-was-off-in-production-and-the-fallback-hid-it https://saytu.ai/blog/semantic-retrieval-was-off-in-production-and-the-fallback-hid-it Wed, 23 Sep 2026 02:55:04 GMT Saytu's semantic retrieval was silently off in production while keyword fallback kept answers flowing. A backfill exposed the 401 and the environment-variable bug behind it. The product match was right; the card still made it look wrong https://saytu.ai/blog/the-product-match-was-right-the-card-still-made-it-look-wrong https://saytu.ai/blog/the-product-match-was-right-the-card-still-made-it-look-wrong Tue, 22 Sep 2026 11:00:52 GMT Saytu found the right product variant but showed a misleading main image. We changed the recommendation rule so variant-only matches are made explicit in the reply. Nine tied rows were enough to make the same question retrieve different sources https://saytu.ai/blog/nine-tied-rows-were-enough-to-make-the-same-question-retrieve-different-sources https://saytu.ai/blog/nine-tied-rows-were-enough-to-make-the-same-question-retrieve-different-sources Mon, 21 Sep 2026 18:07:52 GMT Nine tied rows made identical retrieval calls return different sources. Why the candidate window and the final ranking both needed a stable order before any quality measurement was worth running. Instant Avatar redraws the moving parts instead of rebuilding the whole character https://saytu.ai/blog/instant-avatar-redraws-the-moving-parts-instead-of-rebuilding-the-whole-character https://saytu.ai/blog/instant-avatar-redraws-the-moving-parts-instead-of-rebuilding-the-whole-character Mon, 21 Sep 2026 16:43:14 GMT How one illustration becomes a blinking, speaking avatar: three marked regions are redrawn and the rest of the artwork stays the bytes the merchant uploaded. Three ceilings a character runs into, and none of them say so https://saytu.ai/blog/three-ceilings-a-character-runs-into https://saytu.ai/blog/three-ceilings-a-character-runs-into Mon, 21 Sep 2026 16:43:13 GMT Three browser limits a Live2D renderer runs into on a storefront, what each one costs, and why none of them report themselves where they happen. The model only gets to paint inside three ellipses https://saytu.ai/blog/the-model-only-paints-inside-three-ellipses https://saytu.ai/blog/the-model-only-paints-inside-three-ellipses Mon, 21 Sep 2026 10:12:20 GMT How six avatar states are generated from one illustration without the character drifting, and the single step in the pipeline that gives that guarantee up. The metafields that crowded out the shop policies https://saytu.ai/blog/the-metafields-that-crowded-out-the-policies https://saytu.ai/blog/the-metafields-that-crowded-out-the-policies Mon, 21 Sep 2026 06:00:18 GMT Row volume is not importance. Why retrieval candidates need per-source quotas, what the numbers are, and the ceiling still left in the code on purpose. What verified means when a source is switched off https://saytu.ai/blog/what-verified-means-when-a-source-is-switched-off https://saytu.ai/blog/what-verified-means-when-a-source-is-switched-off Mon, 21 Sep 2026 06:00:15 GMT Turning a knowledge source off has to mean the assistant stops using it. Why row status could not carry that, and what verified actually claims. Four records, and what the assistant is actually shown https://saytu.ai/blog/four-records-and-what-the-assistant-is-shown https://saytu.ai/blog/four-records-and-what-the-assistant-is-shown Mon, 21 Sep 2026 06:00:12 GMT How a storefront assistant picks which parts of a merchant's own writing reach the model: four records of 520 characters, a cosine floor at 0.25, and two lanes that fail differently. We published 108 articles nobody read https://saytu.ai/blog/we-published-108-articles-nobody-read https://saytu.ai/blog/we-published-108-articles-nobody-read Mon, 21 Sep 2026 06:00:10 GMT 108 generated articles reached our blog through an API that required only a title. What the database says about them, and the gate that now stands in the way. A draft should never take the live answer down https://saytu.ai/blog/a-draft-should-never-take-the-live-answer-down https://saytu.ai/blog/a-draft-should-never-take-the-live-answer-down Mon, 21 Sep 2026 06:00:09 GMT A pointer that outlived its assumption turned drafting an FAQ into an unpublish. What the status clause in the filter is really doing. When every score ties, the ranking is a coin flip https://saytu.ai/blog/when-every-score-ties-the-ranking-is-a-coin-flip https://saytu.ai/blog/when-every-score-ties-the-ranking-is-a-coin-flip Mon, 21 Sep 2026 06:00:08 GMT Ties in keyword retrieval made the same question return different material. Why the tiebreak is identity rather than quality, and why it had to go in twice. A turn id is whatever the client says it is https://saytu.ai/blog/a-turn-id-is-whatever-the-client-says-it-is https://saytu.ai/blog/a-turn-id-is-whatever-the-client-says-it-is Mon, 21 Sep 2026 06:00:07 GMT Why a client-supplied turn id can never be a key on its own, and how a lookup one field too short leaked a reply between two stores. Three windows, and what the assistant is allowed to keep https://saytu.ai/blog/three-windows-and-what-the-assistant-keeps https://saytu.ai/blog/three-windows-and-what-the-assistant-keeps Mon, 21 Sep 2026 05:47:25 GMT What a storefront assistant remembers about a shopper between visits, exactly: twenty-four hours of transcript, thirty days of subjects, no message text in either. Uploading your own Live2D model, and what actually stops you https://saytu.ai/blog/uploading-your-own-live2d-model-and-what-stops-you https://saytu.ai/blog/uploading-your-own-live2d-model-and-what-stops-you Mon, 21 Sep 2026 04:49:24 GMT The moc3 version ceiling, a 10 MB body buffer that truncates instead of erroring, and a promise that never settles: what really breaks Live2D model upload.