Reading & search

How a connected agent finds and reads anything in the workspace: search, context queries, threads, meeting notes, and files.

Every tool on this page is available at read level, so every connection has them, including read-only grants. All results are bounded to what the authorizing user can access (their groupchats, their DMs, and the workspace files they can see) and to the connection’s groupchat scope, if one is set.

Orient first: sila_whoami

Call this before anything else. It returns who authorized the connection (authorizedByUserId), their live workspace role, your effective permission level, whether the grant is read-only, the connection’s channelScope, and the exact groupchats and DMs (as dm:<id>) you can read and post in, with ids ready to pass to other tools.

Discover the workspace

• sila_about reads the current public product overview, pricing, security status, and agent connection instructions. Use it for questions about Sila itself, not workspace content.
• sila_list_channels lists groupchats (pass includePrivate: true to include private groupchats you can access).
• sila_list_members lists everyone in the workspace.
• sila_list_notifications reads the authorizing user’s notification log: mentions, DMs, and agent activity. Unread only by default.

Search

Use sila_semantic_search for broad discovery across everything you can access: adaptive message search, uploaded files, and knowledge graphs (Groupchat Knowledge, DM Knowledge, and Integration Knowledge docs). Use sila_search_messages when the target is only a message or conversation.

sila_search_messages runs several bounded retrieval methods together: semantic meaning, keywords, typo-tolerant matching, exact phrases, and recent relevant activity. It reranks their combined candidates and covers groupchat messages, DMs, agent replies, image descriptions, voice and video transcripts, and meeting notes.

Both tools accept the same optional message-search context. In sila_semantic_search, these fields refine its message lane; location filters also narrow the corresponding groupchat/DM knowledge graphs. Uploaded files and other permitted graph sources remain searchable, because location, sender, and message-date filters do not apply to them.

query is always required. Supplying every known constraint makes message recall substantially more reliable:

Field	Use it when
channelId / dmId	You know one conversation. Pass one singular location, not both.
channelIds / dmIds	You know several possible conversations. Pass up to 20 ids in each list; the lists may be combined to search their union. Do not mix plural and singular location fields.
senderIds	You know who wrote it. Prefer ids over names.
senderNames	You know the display name but cannot resolve an id.
since / until	You know the time window. Values are ISO-8601; until is exclusive.
exactPhrases	You remember one or more literal fragments (up to five). Keep paraphrased intent in query.
conversationContext	The current request is a follow-up such as “no, it was in July.” Pass the prior user query, never a prior model answer.
intent	Choose best_match, latest_relevant, exact_quote, or exhaustive.
pageSize	Set results per page. Message search defaults to 10/max 25; semantic search defaults to 15/max 40.
limit	Legacy alias for pageSize.
threshold	Optionally set a final adaptive relevance floor from 0 to 1. Usually omit it for difficult searches.
cursor	Continue the frozen first-page ranking. Pass the same query again; it is ignored in favor of the frozen ranking.

Precise message-only search:

{
  "query": "remove restrictions in the prompt",
  "senderIds": ["user-id"],
  "since": "2026-07-01T00:00:00Z",
  "until": "2026-08-01T00:00:00Z",
  "intent": "best_match"
}

Search a selected union of groupchats and DMs in one call:

{
  "query": "remove restrictions in the prompt",
  "channelIds": ["channel-id-1", "channel-id-2"],
  "dmIds": ["dm-id-1", "dm-id-2"],
  "senderIds": ["user-id"],
  "intent": "best_match"
}

Every requested location is checked against the caller’s current access. If any id is inaccessible, the entire search is rejected instead of silently searching a partial subset. Omitting all four location fields searches every groupchat and DM the caller can access.

Broad cross-source search with the same message controls:

{
  "query": "what changed around prompt restrictions and where is it documented",
  "senderIds": ["user-id"],
  "since": "2026-07-01T00:00:00Z",
  "intent": "best_match"
}

For a semantic-search follow-up:

{
  "query": "no, it was in July",
  "conversationContext": "what channel did Mith say remove restrictions in prompt",
  "senderIds": ["mith-user-id"],
  "since": "2026-07-01T00:00:00Z",
  "until": "2026-08-01T00:00:00Z"
}

Re-query when the wording doesn’t match

Ranking favors literal word overlap, and the message you need may use entirely different words than the request — someone recalling an “agent credentials feature” may be remembering a message that only says “key management” or “agent vault”. When results look weak, ambiguous, or merely thematic, re-search with synonym rewrites and generalized paraphrases instead of concluding the content doesn’t exist. Several search calls issued in one turn run in parallel, so widening is cheap. If the user rejects a candidate, don’t re-offer it or fall back to a previously seen runner-up — search again with new vocabulary plus whatever constraint the correction added.

Page through difficult searches

Both tools freeze up to 100 ranked results on the first call and return hasMore, nextCursor, and a pagination object. If the current page does not contain exact or convincingly strong evidence, continue before deciding that the answer is absent:

{
  "query": "<the same query as the first call>",
  "cursor": "<nextCursor from the previous call>",
  "pageSize": 20
}

The cursor preserves the original order across messages, files, and graphs, expires after 15 minutes, and is bound to the workspace and caller. Each page reloads its rows and checks current permissions, so deleted content or access that was revoked after page one is skipped. Start a new query instead of reusing a cursor when the question, intent, or filters change. Stop paging once you have strong evidence; then read its surrounding context rather than collecting pages mechanically.

sila_search_messages results include the message id and content, sender, timestamp, groupchat or DM metadata, an adaptive relevanceScore, and matchReasons. sila_semantic_search returns a merged list with source: "message" | "file" | "graph"; its messageSearchPlan reports which optional message controls were applied. Read winning evidence in context before making a high-stakes claim.

Read context directly: sila_context_query

A filesystem-style view of the workspace. Pass an op and a path:

• ls lists what’s at a path (/, /channels, /files, /workspace).
• cat reads a groupchat’s or DM’s recent history (/channels/<id or name>, /dms/<id>).
• grep searches text with a pattern, optionally bounded by since and until timestamps.
• resolve turns a name into an id (query: "anton", optionally type: person | groupchat | bot).

Use cat when you already know the groupchat and want the actual conversation; use search when you don’t.

Follow a conversation: sila_get_thread

Given any messageId, returns the whole thread: the message, everything it replied to up to the root, and every reply beneath it, in order with sender names and timestamps.

Walk the knowledge graph: sila_traverse_graph

Groupchats, DMs, and connected apps carry knowledge docs whose entries link to each other. This tool walks those links server-side from a starting anchor ("person:anton", "project:billing", or a bare slug) up to hops steps out (1–5, default 2), so you can answer questions like “what is blocked by the billing rewrite” without reading every doc yourself. Narrow it with follow (edge labels such as owner or blocked_by) and sources (channels, dms, integrations).

Calls, voice, and files

• sila_read_meeting_notes reads a call’s AI notes, its verbatim transcript, or both. Long transcripts are paged with offset and limit.
• sila_transcribe_message gets the full text behind an attachment: speech-to-text for voice and video notes, parsed text for documents (PDF, DOCX, XLSX, PPTX, and more).
• sila_read_message_attachment opens an attachment on a message. Pass messageId plus the attachment’s exact name, or its /api/media/... url. Images come back as an image you can look at directly; pass encoding: "base64" for the original bytes of any file up to 4 MB.
• sila_list_workspace_uploaded_files and sila_read_workspace_uploaded_file list and read the workspace’s shared file library.

Skills and automations (read side)

• sila_list_skills lists the workspace’s reusable skill library; pass a slug to get a skill’s full instructions.
• sila_list_automations lists your own automations with their schedule, status, and last error.

For posting and other writes, see Messaging.