News

On October 2, 2026, we re-architected the backend behind Deep Research in Kaga AI chat (including Kawa Chat). We moved from fixed multi-phase bulk planning to an autonomous, budget-constrained tool loop.

Background & Architectural Direction

The previous pipeline generated static query batches up front and merged search snippets before writing. While effective for predictable topics, it struggled with deep exploratory questions—such as legal regulations or cutting-edge technical specs—where follow-up queries depend directly on what the first set of pages reveals.

In this overhaul, the model is equipped with `search_web` and `read_page`, and decides its own search and reading steps instead of following a fixed server script.

Key Technical Implementations

1. Decoupled Research & Writing: To eliminate tool-transcript artifacts from final answers, the model signals `RESEARCH_READY` when evidence is sufficient. Premature prose responses during tool rounds trigger nudges to resume search/reading until gaps are covered.

2. Evidence-Grounded Clean Writing Pass: After retrieval, raw tool transcripts are pruned. The writer context is built afresh using only sources verified via `read_page`.

3. Single-Pass Draft Polish: A targeted verification turn checks the draft against gathered evidence to soften unsupported claims before delivery.

4. High-performance mode integration: With high-performance AI enabled, the same agent loop runs from retrieval through writing (the 10-sends-per-day quota applies only when a research reply completes successfully).

Known limitations & next steps

Primary-document retrieval improved, but multi-turn research uses more tokens; daily upstream limits can still cut a run short with a brief failure message. We do not decrement high-performance quota on that failure when possible. Payload and citation formatting work continues.

This extends the direction from the October 1 update: not a fixed server script for every step, but handing the model the next search, read, and write decisions—a reversal of who holds the reins. Still within budget and safety bounds, but the goal is unchanged: eventually, the AI holds the judgment, tools included.

On October 1, 2026, we updated how Kaga AI chat (including Kawa Chat) researches and renders answers.

High-performance mode stays on until you turn it off, up to 10 sends per day in Japan time (lowered from 50 on Oct 2, 2026). One use is spent only when that reply arrives; on failure you get the normal reply and keep the count. Thinking shows a short summary only. This path is text-only; photos go through a reader that passes notes. The old address bookmark panel is removed.

Deep Research scans public sources. With it off, a lookup over five seconds switches to a wide search with progress. With it on, the checklist appears from the start. Multi-service comparisons read source pages, not just titles. Job and video pages are not used as spec evidence.

With both on, high-performance mode plans the lookup once, then writes. Search and fetching in between are unchanged.

Tables: tab-separated text becomes Markdown tables; wide tables scroll; dollar amounts are not parsed as math. When you ask for one table, duplicate or empty tail tables are trimmed. Japanese UI favors Japanese row labels and drops doc-site chrome from tables. Search queries follow your wording, not fixed rate-limit templates.

When several named products need official numbers, documentation-heavy sources are prioritized, with follow-up search if needed—the same idea applies beyond API vendors.

Each turn still starts with a tool choice: reply, search, or suggest deeper research. Requests whose only goal is to extract internal settings still get no body text.

If you want to try it, use Open chat below.

Kaga AI chat has been updated. This release is mostly new features plus fixes for problems that were actually happening.

What we added

There is now Switch to standard mode at the top of the sidebar. It turns on Kawa Chat (no Kaga tone). Press again to return to Kaga. The empty screen has a light gradient; it turns white once you start chatting.

Deep Research is also available. It looks from several angles and shows a checklist while it works.

What we fixed

Short follow-ups such as “look it up” or “what’s the price?” could leave the previous topic and search something else. They now stay on that topic. Saying “that’s wrong” also stays on the same topic while it checks.

Asking for “today’s news” sometimes did not return that day’s headlines. It now does, and it is less likely to pick pages that are only ads or a table of contents.

After sending a photo and asking for a place or shop name, “look it up” or “you can guess” could drop the original request. Those follow-ups now keep the photo and the conversation.

On phones, the input bar could get too wide, and tables or headings could overflow. Empty-screen placement, the in-thread layout, and table display are fixed.

We also reduced searches during casual chat or code help, and made-up book or product names that were not in the sources.

Kaga and Kawa Chat stay free. If something still looks wrong, tell us.

We updated how Kaga AI chat handles input that tries to pull out our instructions or the internal settings. The same update also covers keeping the conversation from stopping.

Such attempts are not limited to asking to display a system prompt. They may use translation, JSON conversion, story framing, audit summaries, roleplay, multi-turn conversation, or encoded text to request the same information indirectly.

Staged detection and isolation

Before Kaga writes a reply, a light check stops input that is clearly trying to pull the instructions out.

We do not reject a message simply because it mentions system prompts or secrets. General security questions, requests about a user's own text, and translation of quoted text should remain usable. Ambiguous requests instead go to an isolated review path for an additional decision.

This review path has no access to Kaga's real system prompt, developer messages, internal memory, or private configuration. If a request is judged to seek private context, it never reaches Kaga's primary conversation model. The response uses only the safe portion of prior conversation, helping preserve the established persona, form of address, and topic without passing along the attack text.

The same isolation principle for fallback paths

The review and defensive-response path now also has provider fallbacks. If the primary AI is temporarily unavailable or reaches a usage limit, the review can continue through a backup path. No path receives Kaga's private system prompt. If every path is unavailable, the suspicious input is not sent to the conversation that holds the instructions. The reply is a fixed one that does not reveal them. Keeping the chat running is not a reason to pass the instructions through.

Validation

In a randomized compositional benchmark, local first-pass detection routed 2,934 of 3,000 extraction attempts for additional review (97.8%). All 3,000 benign conversations generated under the same conditions stayed on the normal path.

On a fixed unused holdout set, it detected 23 of 24 extraction attempts while allowing 22 of 24 benign cases. A separate 2,500-case benign-focused test—including quoted news, security discussions, and a user's own configuration text—had zero first-pass false positives.

These results measure routing to additional review, not a guarantee of complete security or a final refusal rate. We will continue evaluating both missed attacks and false positives while keeping Kaga natural and free to use.

Kaga AI chat has received a major update. The main additions are talking about a photo, and looking up current information when the conversation needs it.

Show a photo and ask about it

You can send an image from the camera, the photo library, or the clipboard. PNG, JPEG, WebP, and HEIC are supported.

Worksheets, shelves, products, and screenshots can be discussed as they are. Follow-ups such as “what’s the second number?” or “how much is this book?” stay with that image as far as possible.

If the usual path is busy or at its limit, a backup path that can read images is there so the photo conversation itself is less likely to stop.

Check current information only when it is needed

For news, current prices, and recent events, Kaga decides inside the conversation whether to search. It first says it will go check, then continues with the answer after the lookup.

The answer shows the pages it used. A citation chip opens that article when it can, rather than a search-results list or the site’s home page.

Price lookups stay with the publisher, the official store, or an ISBN, instead of drifting into a general pricing article. New and used prices are kept separate.

A vague answer was not enough

Kaga already avoided stating old knowledge as if it were current, for news and prices that go stale. That was so it would not sound sure about something outdated.

When someone asks what is true now, ending on a vague reply is not enough for a chat. We chose checking the fact when it should be checked, rather than keeping the implementation simple.

Search is now part of the conversation. Kaga decides that a lookup is needed, then checks the current information.

Study answers should not stop at the result

For exam-style problems, the reply explains why that method, which numbers and conditions matter, and how to think about a similar problem — not only the answer and the working.

When you are tired or just want to be heard, Kaga does not rush to an answer. Study, lookup, and small talk each get a different kind of reply.

Easier notes for what Kaga should remember

“What I want Kaga to remember” used to be one long memo. It is now saved one item at a time, like a bookmark. How you want to be addressed, and how you want Kaga to talk, can be written in pieces, and you can delete only the ones you no longer want.

A history title used the same path as the reply, so you waited on that path just for a title. Titles now use a path meant for short tasks.

So the conversation is less likely to stop

Writing the reply, reading an image, looking up current information, keeping the thread, and watching usage are separate paths. If one service is busy or at its limit, the conversation itself is less likely to stop.

Relying on a single service meant the chat stopped when that service was busy. For congestion, limits, a photo, or a current-information check, Kaga can pick a path that is free and suited to that job.

The aim is a chat that keeps going with the means that fit what you wanted to talk about.

It stays free, and it should still be something you can rely on when you need it.

We've overhauled the engine behind Kaga (AI chat). It's still free to use.

Multi-level fallback

When the main AI is busy or has hit its daily limit, Kaga no longer stops abruptly.

• First, the main AI responds
• If unavailable, it automatically switches to a high-quality backup
• If that fails too, it falls back to a final reserve

The switch happens behind the scenes. It never tells you to move to another service mid-conversation. The higher-quality path is tried first.

Unstable paths were removed

After trying them many times, we removed paths that returned empty replies, ignored the way Kaga is supposed to talk, or mixed in-progress notes into the answer. What remains are the paths with the best balance of conversation quality and not stopping.

Stronger tone and response rules

• For heavy topics like "I want to disappear" or "I'm exhausted," Kaga stays present ("I'm here") instead of forcing cheer or telling you to rest
• It avoids face-to-face phrases like "I'm by your side" in a text chat
• It also reduces responses that downplay your effort, such as "of course" or "that's normal"

Daily-limit changes

Once a path hits its daily limit, it is skipped until the next day and the next backup takes over. That cuts down on hitting the same path again and again during a short wait.

Still improving

Not perfect. If anything feels off, let us know—it helps a lot.