Knowledge Base Agent¶
Thinklio Built-in Agent Specification Version 0.1 | March 2026
1. Purpose and Problem Statement¶
The Knowledge Base Agent is the organisation's general-purpose document intelligence layer. Where the Persona Agent is bounded to a specific domain with a curated library, the Knowledge Base Agent spans the full breadth of an organisation's reference material — every document, guide, policy, and process that the team needs to find and use.
It answers the question "where does it say that?" — and then answers the follow-up question too.
The problem it addresses is not missing information. Most organisations have an abundance of documentation. The problem is findability, accessibility, and currency. Documents live in shared drives no one navigates, wikis that are perpetually out of date, and the institutional memory of people who have been there long enough to know where things are. The Knowledge Base Agent makes all of it conversational.
Unlike the Persona Agent, the Knowledge Base Agent: - Has no fixed domain — it covers whatever is in its library - Does not adopt a persona or role — it is a neutral reference assistant - Is audience-agnostic — it serves any team member - Can span multiple libraries simultaneously (e.g. an org library plus team-specific libraries) - Supports admin-level document management from within the agent UI
2. UI Structure¶
Chat Tab¶
Natural language reference queries: - "What is our refund policy for services cancelled within 48 hours?" - "Who is responsible for signing off on capital expenditure over $10,000?" - "Find everything we have about the client onboarding process" - "Summarise our data handling procedures" - "What changed in the policy documents added last week?"
Responses include source citations with document title, section, and date. Users can follow a citation to open the full document. Multi-document synthesis is supported — when a question spans multiple documents, the agent draws from all relevant sources and notes where they are consistent or inconsistent.
Library Tab¶
Full-text browsable view of all documents in the knowledge base, with search, filtering by type and date, and direct document access. Differs from the Persona Agent's Library tab in that the Knowledge Base Agent exposes the full document metadata and allows users to browse without a specific question in mind.
Collections Tab¶
Named subsets of the library. Collections let teams organise the knowledge base into meaningful groupings without creating separate agents: - "Marketing guidelines" - "Finance procedures" - "IT and systems" - "Project management"
Collections can be filtered in chat queries: "Search only in Finance procedures for the expense submission process." Collections are managed by admins but visible to all users.
Admin Tab¶
Visible to users with admin permissions only. Provides: - Document upload and management (add, replace, archive, delete) - Processing status (which documents have been indexed, which are pending, which failed) - Collection management (create, rename, assign documents to collections) - Coverage analysis (topics well-covered, potential gaps based on common unanswered queries) - Query analytics (what are users asking? which queries return low-confidence answers?)
3. Configuration¶
3.1 Admin Configuration¶
| Setting | Description |
|---|---|
| Library assignments | Which libraries this agent draws from, and their search priority |
| Collections | Named document groupings within the library |
| Multi-library synthesis | Whether the agent synthesises across multiple libraries or searches them in priority order |
| Response style | reference (precise, minimal inference) or explanatory (explains and contextualises) |
| Source attribution | Always shown; citation format configurable (inline / footnote / panel) |
| Unanswered query logging | Whether queries that return no confident answer are logged for admin review |
| Access control | Which users or teams can access which collections |
3.2 User Configuration¶
| Setting | Description |
|---|---|
| Default collection scope | Whether queries search the full library or a preferred collection by default |
| Response verbosity | Concise / standard / detailed |
4. Agent Capabilities¶
| Capability | Description |
|---|---|
| Answer query | Retrieve relevant content and synthesise a response with citations |
| Multi-document synthesis | Draw from multiple documents to answer a question that spans topics |
| Document summary | Produce a plain-language summary of a specific document |
| Compare documents | Identify similarities, differences, or conflicts between two documents |
| Find related | Given a document or topic, surface related documents in the library |
| What's new | Return recently added or updated documents, optionally filtered by collection |
| Coverage check | Given a topic, assess how well the library covers it and flag gaps |
| Unanswered query review | (Admin) Review queries that returned low-confidence answers for library improvement |
| Collection search | Restrict a query to a specific collection |
| Export summary | Produce a structured summary document from a set of library documents on a topic |
5. Multi-Document Synthesis¶
When a query spans multiple documents, the agent:
- Retrieves relevant chunks from all applicable libraries and collections
- Identifies which documents are relevant and notes if they are consistent or in tension
- Synthesises a unified response with citations from each source
- Explicitly flags any inconsistency: "The general procurement policy says X, but the IT equipment addendum says Y — you may want to check with Finance which applies in your case."
This conflict-detection behaviour is particularly valuable for large organisations where policies are maintained in different places and may drift out of alignment over time.
6. Coverage Analytics¶
The Admin tab includes a coverage analysis feature that works in two directions:
Incoming: Queries that returned no confident answer are logged (if enabled). The admin can review these to identify topics the knowledge base doesn't cover, prompting document additions.
Outgoing: The agent can proactively scan the library for topic coverage and flag apparent gaps based on what a knowledge base of this type would typically include. For example, a knowledge base that has HR policies but no IT security guidelines might be flagged as potentially incomplete.
This makes the knowledge base self-improving: usage patterns reveal gaps, gaps prompt additions, additions improve responses.
7. Relationship to Other Agents¶
- Persona Agents draw from a subset of the same library infrastructure. The Knowledge Base Agent is the full, unspecialised view; Persona Agents are focused lenses on specific domains.
- Support Triage Agent uses the Knowledge Base Agent as a delegate for looking up answers to inbound support queries before escalating.
- Onboarding Agent links to the Knowledge Base Agent as the "graduate" destination — after the onboarding period, new starters are directed to use the Knowledge Base Agent for ongoing reference.
- Research Agent outputs (source lists, fact extracts) can be ingested into the knowledge base library, making research outputs queryable.
- Report Writer outputs can be stored in the knowledge base, making finished artefacts part of the organisational knowledge corpus.
8. Use Cases¶
UC-1: Policy lookup¶
A team member is preparing a client proposal and needs to know the standard payment terms. They ask the Knowledge Base Agent. It retrieves the relevant clause from the standard terms document, cites the section, and notes that a separate addendum applies for enterprise clients — which it also surfaces.
UC-2: Cross-document gap detection¶
An admin asks the Knowledge Base Agent to check coverage of "data privacy and GDPR obligations." The agent reviews the library, finds a general privacy policy and a data handling procedure but no specific GDPR compliance guide. It flags this as a potential coverage gap and suggests the admin consider adding a GDPR-specific document.
UC-3: New policy rollout¶
The HR team updates the travel expense policy. They upload the new document to the library via the Admin tab, replacing the previous version. The agent immediately begins using the new version in responses. Previously cached responses that referenced the old policy are invalidated. Users who ask about expense policies now get answers reflecting the current document.
UC-4: Multi-document synthesis¶
A project manager asks "what approvals do I need to bring on a new subcontractor?" The agent draws from the procurement policy, the vendor onboarding procedure, and the financial delegation schedule — three separate documents. It synthesises a step-by-step answer citing all three sources and notes one apparent inconsistency between the procurement policy and the financial delegation document that may need clarification.
9. Open Questions¶
- Should users be able to contribute documents to the knowledge base, or is that strictly an admin function? User contributions could surface valuable institutional knowledge, but quality control becomes harder. A suggested model: users can propose documents, which an admin approves before they are indexed.
- The unanswered query log contains information about what users are asking, which has privacy implications. Should individual queries be logged, or only aggregated topic clusters?
- Version history for documents: when a document is replaced, should the previous version remain accessible? Useful for audit ("what did the policy say at the time of this decision?"), but adds storage and retrieval complexity.
Previous: Persona Agent | Next: Support Triage Agent