Just Talk to Tasukiba — Semantic Search Chat in Your AI Operations Secretary

“Wasn’t there a retrospective last month about resource shortage?”

This series is walking through Tasukiba Knowledge Relay’s three unique features, one by one.

1. Suggestion engine (released) — I covered in B-2, past assets get pushed to the screen at project create / issue file time
2. Chat semantic search (released, today’s topic) — pull past assets via natural language
3. The “Why?” feature (released, tomorrow’s post) — suggestions get a “why this is related” explanation

Today is feature 2, chat semantic search. It shipped at release — available anytime from the owl FAB at the bottom-right of every dashboard page. Here’s the experience design behind it.

👉 Tasukiba product page

From structured form to natural language

B-2’s suggestion engine takes a structured form as input:

Project name: E-commerce site renewal
Purpose: ...
Background: ...
Scope: ...

The moment that’s saved, the suggestion engine runs and related past assets appear on screen. That’s the Phase 1 experience.

Chat semantic search replaces the entry point with natural-language conversation:

User: "Wasn't there a retrospective last month about resource shortage?"

Tasuki Owl:
  Matching retrospective:

  - 2026-03-15: Sprint 5 retrospective
    > "QA resources were short, test phase slipped"
    > Related risk: R-005 (resource estimation accuracy)
    > Related knowledge: K-012 (criteria for QA automation)

“Fill in a form” becomes “speak to it.” That’s the only surface change. But on the user’s side, it’s a completely different experience.

The same embedding foundation, with a different entrance

This is chat semantic search’s technical identity.

The chat query is also vectorized through Voyage AI into 1,024 dimensions.

As B-2 covered, every knowledge / risk / issue / retrospective / memo / project has a Voyage embedding stored in content_embedding at create/update time.

Chat semantic search just runs the user’s utterance through Voyage, then pgvector finds the closest assets.

Utterance → Voyage → query vector
                       ↓
              Cosine similarity in pgvector
                       ↓
              Top-N retrieved
                       ↓
              Rendered as a chat reply

It looks like a new feature. Under the hood, it’s the same engine as the suggestion feature. What changed is the entrance and the exit.

That’s the design intent behind “don’t build the three features separately.” The embedding foundation built for the suggestion engine gets reused as-is, which kept the marginal cost of adding chat low.

The experience I want to deliver — “ask the moment you wonder”

Forget the architecture for a moment. The experience matters more.

The moment I’m imagining:

You’re looking at the project management screen, and suddenly think, “wait, didn’t we run into something similar on last year’s project?”

In a typical world:

1. Try to remember which memo
2. Type keywords into search
3. Can’t remember, give up

In Tasukiba’s world, the Tasuki Owl is sitting in the bottom-right corner as a chat button. You click and just say, “I think there was similar trouble on last year’s DB migration project.”

A reply comes back. “Matching issue: 2025-09-12, Sprint 3 — ‘Migration script underestimated row count, batch timed out.’ Related risk: R-018 (batch execution time). Related knowledge: K-031 (PostgreSQL copy batch patterns).”

— that “ask the moment you wonder” distance is the experience chat semantic search delivers.

The suggestion engine is a push notification at screen-transition time. Chat semantic search is a companion you can talk to the instant something occurs. Two different surfaces, one shared foundation.

”Chat” and “full-text search” are different things

I don’t want anyone to misunderstand this, so I’ll spell it out.

Tasukiba’s chat semantic search is not keyword-based full-text search. It’s also not a general-purpose LLM chat like ChatGPT.

Kind	Mechanism	Can do	Cannot do
Full-text search	Keyword match	Find exact-match text	Surface variation, concept proximity
General LLM chat	Large language model	General conversation	Ground answers in your past data
Tasukiba chat semantic search	Embed query → similarity with your content_embedding	Respond grounded in your past assets via semantic similarity	Generate info that doesn’t exist in your DB

Tasukiba’s chat is designed to ground every answer in the assets your organization actually accumulated. Hallucination risk is structurally minimized — and that’s what makes it usable as business SaaS, not just a toy.

The Tasuki Owl responds — coupled with mascot design

As M-2 covered, the Tasuki Owl has three master images. One of them, the chat-specific avatar, was designed precisely for this feature.

The owl framed by a speech-bubble outline signals “you can talk to me” at a glance. The character direction from A-5 — “quiet, watching, not pushy” — carries over into the chat UI.

Not “I’ll teach you,” but “let me help you notice.” Not “look here,” but “here, this is sitting on the table.” The mascot’s copy direction becomes the chat reply’s voice directly.

Chat as a thread connecting “three places”

A closing note on the philosophy side.

In A-2, I named five user experiences I want Tasukiba to deliver:

• A. First impression — no friction
• B. Daily use — energy / a place to settle / fresh discovery
• C. When things go wrong — quiet awareness
• D. Long-term — raises success rate
• E. As a personality — a trustworthy companion

Chat semantic search hits especially hard on B’s “a place to settle” and “fresh discovery.”

A lonely PM has someone to talk to sitting beside them. A casual utterance returns records from your past self or team. That’s the temperature one layer beyond Tasukiba’s “business SaaS but also a place” identity.

Tomorrow’s post: the third unique feature, the “Why?” feature — also shipped — where suggestions carry a “why this is related” explanation. LLM re-ranking and the Pro plan differentiation.

Related posts

• The suggestion engine — putting “the past you forgot” back on screen — series part 8, foundation chat sits on
• Post-release roadmap — Tasukiba Phase 2/3 — the post-release roadmap overview
• Three plans, usage-based — pricing to continue — the reasoning behind lightly charging for chat

About Tasukiba

Chat semantic search shipped at release (2026-06-01). Each feature’s experience is on the product page.