How AI Search Works

When you ask an agent a question in Agency Hero, it doesn’t just look for exact keywords. It understands meaning — so “what were we worried about in the last launch?” and “launch risks” can surface the same results. This article explains how that works, what the agent can search, and how to get the best results.

Semantic search vs. keyword search#Copy link

Traditional keyword search matches the words you type, letter-for-letter. If the document says “risk” and you type “concern”, you get nothing.

AI search is different. Agency Hero converts content — documents, meeting summaries, intelligence items, topics — into a numerical representation called an embedding. Embeddings capture meaning, not just words. When you ask a question, the agent converts your query into an embedding too, then finds content whose meaning is closest to yours.

This means:

• “budget concerns” can match content about “financial risk” or “cost overruns”
• “hiring plan” can match notes on “headcount” or “recruiting strategy”
• You don’t need to guess the exact phrasing someone used when they wrote something

The engine uses a similarity score (0–1) to rank results. Items above a threshold (typically 0.55) are returned; items below it are filtered out as irrelevant.

What the agent can search#Copy link

Agency Hero has multiple distinct search backends, each covering a different type of content. The agent picks the right one — or queries several at once — based on what you’re asking for.

1. Documents & knowledge base#Copy link

Tool: search_workspace_knowledge

This covers all uploaded or generated artifacts in your workspace: statements of work, proposals, commercials, objectives, meeting summaries, org-level templates (proposal and proof templates), and any custom documents you’ve uploaded.

Every document is automatically chunked into ~1,000-character segments and embedded, so even long files are fully searchable.

Best for: “Find the SOW for Acme Corp”, “What does our proposal say about pricing?”, “Show me meeting summaries about onboarding.”

2. Intelligence items#Copy link

Tool: search_canonical_intelligence

Intelligence items are structured records extracted from meetings and conversations — things like decisions, risks, questions, commitments, objections, assumptions, blockers, and milestones. They’re confirmed facts about your workspace, not raw discussion.

Best for: “What decisions have we made about the product roadmap?”, “What risks are open?”, “Find any commitments we made to the client.”

3. Discussion topics#Copy link

Tool: search_workspace_topics

Topics are the recurring themes the system extracts and aggregates across your meetings — things like “DevOps infrastructure”, “Pricing strategy”, or “Hiring pipeline”. Each topic has a brief that summarizes all discussions where it appeared and shows how many times it’s come up.

Best for: “What topics are we tracking?”, “Tell me about our DevOps discussions”, “What themes keep coming up in client calls?”

4. Meeting summaries#Copy link

Tool: search_meeting_summaries

This searches the AI-generated summaries produced after each meeting. It’s a fast way to find which meetings touched on a particular subject without reading full transcripts.

Best for: “What meetings discussed the launch timeline?”, “Find past calls where we talked about hiring.”

5. Meeting items#Copy link

Tool: search_meeting_items

This searches the structured items extracted from individual meetings — action items, decisions, risks, and questions captured at the per-meeting level. Useful when you want to find something that came out of a specific meeting rather than across the full intelligence ledger.

Best for: “What action items came out of last week’s call?”, “Find the risk that was flagged in the Acme meeting.”

6. Meeting transcripts#Copy link

Tool: search_transcript

This searches the raw spoken content of your meetings, word-by-word. Useful when you remember someone said something specific but aren’t sure which meeting it was from.

Best for: “Find where someone mentioned the API deadline”, “Which call did we discuss quarterly targets?”

7. Memos#Copy link

Tool: search_memos

Memos are notes created by you or the agent in any workspace. Memo search uses full-text matching across memo titles and body content.

Best for: “Find my memo about the go-to-market plan”, “Search memos for onboarding notes.”

8. Tasks#Copy link

Tools: tasks_list, tasks_get

Tasks are fully searchable by the agent. You can ask the agent to list your tasks, filter by status, find tasks by keyword, or retrieve a specific task by ID. The tasks_list tool accepts a query parameter for keyword matching and a completed flag to include or exclude completed tasks.

Best for: “What tasks do I have open?”, “Find the task about the pricing deck”, “List all overdue tasks in this workspace”, “Show me completed tasks from last week.”

9. Code & repositories#Copy link

Tools: search_code, list_indexed_repos, get_file_content, list_repo_files

If your workspace has connected GitHub repositories, the agent can search across indexed code using semantic similarity, retrieve specific files by path, list repository files, and enumerate all indexed repos. Code search always runs across all indexed repositories — not just the current workspace — because repos may be indexed in a different workspace than the one you’re chatting from.

Best for: “Find the auth middleware implementation”, “Show me the task activity file”, “What repos are indexed?”, “Search for where we handle meeting embeddings.”

10. Published content items#Copy link

Tool: search_content_items

Searches published documentation articles, blog posts, pages, and other content-engine items. Useful for finding related published docs or checking whether an article already exists before creating a new one.

Best for: “Find docs about meeting summaries”, “Is there already an article about workspace topics?”, “Search blog posts about AI agents.”

How vectorization works (the simple version)#Copy link

When content arrives in Agency Hero — a document upload, a meeting summary, a new topic — a background process automatically turns it into embeddings and stores them alongside the original content. You don’t need to do anything for this to happen.

Here’s the flow in plain terms:

1. Content is prepared — The document or summary is serialized into clean text.
2. Text is split into chunks — Long content is broken into ~1,000-character segments with a small overlap between them, so nothing important falls at a boundary.
3. Embeddings are generated — Each chunk is sent to an AI embedding model (OpenAI’s text-embedding-3-small), which returns a list of 1,536 numbers that mathematically represent the meaning of that text.
4. Embeddings are stored — The numbers are saved in a vector database alongside the original text. This is what the agent queries when you search.

When you ask a question, your query goes through the same process in real time (step 3 only), and the system finds stored chunks whose numbers are most similar to yours. Results are returned ranked by similarity.

Why this matters for you: Because the system stores meaning rather than words, you can search naturally. You don’t need to recall exact phrases, and you don’t need to search different synonyms to be thorough.

How the agent decides which tool to use#Copy link

When you send a message, the agent classifies your intent before deciding where to search. There are three main patterns:

State questions#Copy link

“What are our open risks?”, “What topics are we tracking?”

The agent checks the context pack first — a pre-built snapshot of your workspace’s current state. This is faster than searching and works well for high-level overviews.

Find questions#Copy link

“Find the task about the pricing deck”, “Show meetings about onboarding”

The agent goes directly to the most relevant backend. If you’re asking for a task, it queries tasks_list. If you’re asking for a decision, it searches intelligence. If you’re asking for a document, it searches the knowledge base. No need to fan out.

Broad knowledge questions#Copy link

“What do we know about DevOps?”, “Tell me everything about the launch”

The agent fans out across all relevant backends in parallel — topics, intelligence, documents, and meeting summaries — then synthesizes the results into a single answer. This gives you a comprehensive view rather than a single-source guess.

Entity-to-tool routing at a glance#Copy link

Scope: current workspace vs. all workspaces#Copy link

By default, every search is scoped to the workspace you’re currently in. This keeps results focused and relevant.

If you want to search across multiple workspaces, say so explicitly:

• “Search all workspaces for the Acme proposal”
• “What decisions have been made across all my workspaces?”
• “Find mentions of this topic everywhere”
• “What’s on my plate today” (implies a personal cross-workspace view of tasks and meetings)

When you ask for cross-workspace results, the agent queries each workspace in parallel and groups results by workspace name so you can tell where each item came from.

You can also target a specific workspace by name:

• “Search the Vertice Labs workspace for DevOps topics”
• “Find meeting summaries in Acme Corp about the launch”
• “List open tasks in the Content workspace”

Note: Code search always runs across all indexed repositories regardless of scope, because repos may be indexed in a different workspace than the one you’re chatting from.

Tips for better search results#Copy link

Getting great results is mostly about being clear and specific. Here’s what works:

Use natural language, not keyword strings.
Instead of: “pricing SOW Acme”
Try: “What does our statement of work say about pricing for Acme Corp?”

Specify what type of thing you’re looking for.
Saying “find the decision about…” or “what tasks exist for…” or “what risks exist around…” helps the agent route directly to the right backend instead of guessing.

Be specific about the subject.
“What do we know about the Q3 launch timeline concerns?” will get better results than “tell me about Q3”.

If results are sparse, try a related phrase.
Semantic search is flexible, but very niche terminology might benefit from an alternative framing. If “budget overrun” returns nothing, try “cost increase” or “financial risk”.

Ask for cross-workspace results when you need them.
If you suspect something lives in a different workspace, say “search all workspaces” explicitly — the agent won’t expand scope on its own.

Trust broad questions to fan out.
For open-ended questions like “what do we know about X?”, you don’t need to specify the source. The agent will search topics, intelligence, documents, and meeting summaries together and synthesize a complete picture.

For tasks, use direct language.
Try “list my open tasks”, “find the task about the pricing deck”, or “what tasks are overdue?”. The agent filters out completed tasks by default — add “show all” or “include completed” if you want the full history.

What the agent cannot search#Copy link

A few things fall outside AI search at this time:

• Upcoming meetings — Use “What meetings do I have coming up?” to list scheduled meetings via the list_upcoming_meetings tool.
• External integrations — Content in Asana, HubSpot, ClickUp, Teamwork, Salesforce, and similar tools is accessed through integration-specific tools (via the workspace-integration subagent), not workspace search.

Note on tasks: Tasks are searchable. Ask the agent directly — “What tasks do I have?”, “Find the task about the pricing deck”, or “List open tasks” — and it will query your task list immediately.