Search vs Chat
Kaizen gives you two surfaces over the same authorized matching pipeline. They feel different on purpose, and knowing when to reach for each will make you much faster.
Search: find sources fast
Search is non-generative. You type a query, Kaizen retrieves and ranks the best-matching documents, Projects, and prior discussions, and hands them straight back as cards.
Use Search when you want to:
- locate a specific document, policy, passage, Project, or prior discussion,
- see what sources exist on a topic,
- get to an authoritative original quickly and read it yourself.
Chat: get a grounded answer
Chat is conversational and generative. The assistant retrieves relevant sources, reasons over them, and writes an answer with inline citations. It can also use tools — reading a document, running code, producing an artifact, managing a task, or changing an approved platform resource.
Use Chat when you want to:
- synthesize an answer from several sources,
- ask follow-ups and refine,
- have Kaizen do something, not just point you at a page.
When to use which
| You want to… | Use |
|---|---|
| Find the exact doc or passage | Search |
| Reopen where a topic was discussed | Search or Chat |
| Understand or summarize across sources | Chat |
| Move fast and read the original yourself | Search |
| Ask follow-up questions in context | Chat |
| Have Kaizen run code or take an action | Chat |
One pipeline underneath
Both surfaces call the same server-owned matching pipeline. That is what keeps them consistent: a document that Search ranks first is the same document Chat retrieves, under the same authorization, in the same order.
Surface behavior and answer generation differ. The authority to match does not.
Shared history discovery
Every eligible chat agent receives the same conversation_history_search
capability from the common runtime. It runs as the person who initiated the
turn, rechecks current direct and Project room access, and returns bounded
excerpts with safe links. You can ask "where did we discuss the database
migration?" without switching surfaces.
History results carry navigation authority, not evidence authority. They can help an agent point you at a prior chat or Project, but they do not enter the factual citation ledger or silently become grounding for a new answer. An agent must retrieve an eligible document source separately before citing it as evidence.
The capability is shared infrastructure, not a special feature of one persona, so its scope, timeout, result limits, and access checks apply uniformly to every eligible agent.
Default agents and modes
Both surfaces are powered by agents. Each surface has its own default
agent, and every agent declares which surface it belongs to — its mode,
search or chat.
- The Search default is tuned for retrieval: how many results, and which knowledge to consider.
- The Chat default is tuned for a tight, well-cited answer, and searches more sparingly per turn.
When routing is enabled, Chat's Default agent can recognize a high-confidence gap covered by its reviewed specialist roster. A public technical comparison, for example, is handed to Web Researcher as a visible child run; Default then builds the final response from that result. A low-confidence or tied match stays on Default. Routing requires an explicit reviewed tag match, so ordinary phrases in a direct-answer request do not hand the turn to the only available specialist. The composer can turn off a handoff for one prompt.
Because agents are mode-scoped, a Search agent never appears as a Chat option and vice versa. The matching mechanism stays shared, so the two surfaces can behave differently without creating competing document, chat, or Project search authorities.
Compression does not change evidence
Tool-result compression does not alter first-party search, conversation-history search, document retrieval, or any other source-bearing evidence. Those model-bound payloads bypass compression byte-for-byte.
Query construction, authorization filtering, ranking, retrieved source IDs, answer facts, and citations therefore follow the same path whether a member's compression setting is on or off. See Context management & compaction.