Notes from the Saytu team

What we learn building a Live2D character that answers preorder, sizing and authenticity questions on anime storefronts.

Latest

Build logOctober 1, 20265 min read

The FAQ draft step calls no model

Our FAQ drafting step deliberately calls no model. It copies answers only from evidence we already have, preserves the merchant's own words, and leaves the answer blank when nothing trustworthy exists instead of inventing store policy.

More from the team

Semantic search was off in production. Keyword search made it look healthy.

Semantic retrieval was broken in production, but keyword fallback kept answers flowing and hid the failure. We fixed the credential, backfilled current embeddings, and made keyword-only mode visible instead of treating any answer as proof of retrieval health.

Build logSeptember 30, 20265 min read

We let embeddings choose the product, but never the price or stock

We use semantic retrieval to decide which product a shopper means, but we never trust the index for changing commercial facts. Price, stock and availability are read live from Shopify before they become part of the answer.

Build logSeptember 29, 20265 min read

One handoff flag was answering two different questions

We had one handoff flag doing two jobs: deciding whether the AI could speak and whether a human still owed the shopper a reply. Splitting AI Paused from Needs Human turned a fallback into a real lifecycle.

Build logSeptember 28, 20265 min read

The model understood the shopper. The catalogue did not.

A shopper asked naturally for a product, the model understood, but the catalogue never opened because an intent gate missed the phrasing. We changed the routing rule rather than the model, because natural-language support has to include everything around the LLM too.

Build logSeptember 25, 20265 min read

We kept the subject for 30 days instead of keeping the conversation

A shopper returned the next morning with the same session but without yesterday's transcript. Instead of retaining more conversation, we kept only the subject of a successful product search for the session's 30-day lifetime.

Build logSeptember 24, 20265 min read

The answer was right. The black shirt underneath it made it wrong.

We matched a shopper's request for white to a product that really had a white variant, then showed its black main photograph. The reply was technically right and the experience was wrong. We changed the rule so product cards have to support the answer a shopper actually sees.

Build logSeptember 24, 20265 min read