Context and grounded answers

Context tells Usable Chat which maintained knowledge and supporting material are relevant to the conversation. Good context improves retrieval and reduces ambiguity, but it does not make every generated statement automatically correct.

Treat Chat as a reader and reasoner over sources: select the right boundary, ask it to retrieve evidence, and verify important claims before acting.

Context you can use

Depending on the current surface and configuration, context can include:

• workspaces;
• memory fragments;
• collections;
• project context;
• selected files and current-turn images;
• runtime context supplied by an embed host.

A workspace defines the main permission and knowledge boundary. Collections remain workspace-scoped, so include the workspace when collection context must be resolved reliably.

Add context in the web app

Use the context controls shown in Chat to select accessible workspaces, fragments, collections, or project material. Another Usable page can also open Chat with context already staged.

Released first-party links support comma-separated query parameters such as:

/chat?workspaces=<workspace-id>&fragments=<fragment-id>

The link stages pending context for the first turn. It does not bypass permissions. After a new conversation is created, Chat removes consumed workspace and fragment parameters so refreshing does not repeatedly inject the same context.

Embed mode uses its configured default context and parent bridge instead of the first-party URL behavior.

Scope narrowly enough

More context is not always better. A broad collection of unrelated workspaces can increase noise and make it harder to identify the authoritative source.

Prefer this order:

1. Select the workspace that owns the answer.
2. Add a specific fragment or collection when known.
3. Add files only when they contain necessary evidence.
4. Add web search only when current external information is required.

If two workspaces disagree, ask Chat to show the conflict rather than silently combining them.

Ask grounded questions

State both the task and the evidence rule:

Answer from the selected workspace. Fetch the complete relevant fragments, name the sources you relied on, and label anything that is not confirmed by those sources.

For a time-sensitive question, include the relevant date or release. For a procedure, ask for prerequisites, steps, verification, and failure cases. For a comparison, ask Chat to keep source-backed facts separate from recommendations.

Workspace sources and web citations

Usable context and web search are different source classes:

• Workspace context uses knowledge your account can access in Usable.
• Web search retrieves current external results when enabled and allowed.

The UI may show source blocks with titles, links, summaries, or scores. A citation proves that a source was retrieved; it does not prove that every sentence correctly represents it.

Ask Chat to distinguish workspace evidence from external web evidence. Do not let a current web result silently override maintained internal policy.

Verify an important answer

1. Open the cited fragment, file, or web page.
2. Read the complete relevant passage, not only the source snippet.
3. Confirm its workspace, author, date, status, and release context.
4. Check whether the answer mixed sources with different audiences or versions.
5. Re-run the question with a narrower scope if evidence is missing.
6. Ask a responsible person to review decisions with legal, financial, security, or operational impact.

For Usable fragments, clear titles, summaries, tags, and focused content improve what Chat can discover. Retrieval quality depends partly on the quality of the knowledge base.

Permission boundaries

Context is always limited by access. A URL, fragment ID, collection ID, or embed setting cannot grant access that the signed-in account or credential does not have.

If context disappears or a source cannot be read:

• verify the signed-in account;
• verify the workspace role or public subscription;
• include the workspace for a collection;
• check whether the fragment is active and accessible;
• check whether an embed bearer token has expired.

Do not broaden workspace access merely to make one answer work. Move public-safe knowledge to the correct workspace or ask an owner for the minimum appropriate access.

Context in embedded assistants

An embed administrator can set default workspaces, collections, and fragments. A host application can add or replace runtime context after the iframe is ready.

Users should be able to understand the assistant's effective scope. Host-provided page context is useful, but it should not contain secrets or hidden data that the user is not authorized to discuss.

See Configure an embedded assistant and Integrate embedded Chat.

Troubleshooting

The answer is generic

Select a more specific workspace or fragment and ask Chat to cite it. Confirm that the source contains enough detail to answer the question.

A collection is ignored

Include its workspace. Collections are resolved within a workspace boundary.

Citations are missing

Ask for sources explicitly and confirm source-citation visibility is enabled. Some model or tool paths may return an answer without a citation block; verify manually before relying on it.

Web results dominate the answer

Disable web search or instruct Chat to treat workspace material as authoritative and web results as supplementary.

Related pages

• Using Usable Chat
• Memory fragments
• Search and retrieval
• Workspace roles and permissions