A Claude-powered operations agent that turns approved clinic documents, spreadsheets, and procedures into source-linked answers, drafts, and staff actions.
The Knowledge Base AI Assistant is a working internal agent for multi-location healthcare and wellness teams. It searches approved Google Workspace documents and spreadsheets, answers with source references, drafts customer follow-ups, retrieves training procedures, and escalates uncertain requests instead of inventing an answer. The runtime uses the Anthropic Messages API, while controlled actions are exposed through Claude tool use.
Why This Knowledge Base AI Assistant Fits Multi-Location Clinics
Clinic information often lives in shared drives, onboarding documents, scheduling sheets, procedure notes, and informal staff messages. That fragmentation creates different answers between locations and forces managers to repeat the same guidance.
This build creates one governed retrieval layer without replacing the systems already in use. Anthropic’s analysis of 100,000 Claude conversations estimated an average 80% reduction in task time for the analyzed work. McKinsey’s 2025 State of AI survey also found that stronger AI performers are more likely to define when model outputs require human validation.
The assistant follows that operating pattern: bounded sources, citations, review gates, and explicit no-answer behavior.
Knowledge Base AI Assistant Architecture
A staff question enters through the web console or command interface. The retrieval service filters sources by location, department, document status, and user role. Claude receives only the relevant excerpts, then either answers, drafts an approved action, or sends the request to a manager queue.
The connector layer uses the Model Context Protocol for reusable system access. The maintenance package uses Claude Code for connector tests, prompt checks, schema updates, and repetitive deployment tasks.
Choosing the Right Claude Surface
The included workflow audit prevents every task from being forced into the same interface. Staff knowledge retrieval runs through the Claude API, connector development and diagnostics use Claude Code, and one-off document analysis can remain in Claude.ai.
Browser or spreadsheet agents are reserved for tasks that genuinely require visible interface interaction. Stable APIs and MCP connectors are preferred when structured access is available because they are easier to test, permission, and audit.
Core Features
| Feature | Description |
|---|---|
| AI Search Assistant for Company Knowledge | Staff lose time searching folders and asking repeated questions. The assistant retrieves approved passages, names the source file, and returns a direct answer with an escalation path. |
| Custom AI Assistants for Internal Knowledge Retrieval | Different teams need different context. Role profiles separate front-desk procedures, manager playbooks, training material, marketing references, and location-specific instructions. |
| AI Knowledge Assistant for Employees | New and existing staff need consistent guidance during daily work. Employee mode answers procedural questions, creates short training summaries, and blocks restricted manager material. |
| AI Knowledge Assistant for Managers | Managers need visibility rather than another opaque chat window. Manager mode reviews escalations, approves reusable answers, updates source priority, and inspects the query audit trail. |
| Google Workspace Source Sync | Manual uploads become stale. The connector reads approved folders through the Google Drive API and operating data through the Google Sheets API. |
| Draft-and-Review Actions | Repetitive follow-up and content work creates inconsistent wording. The agent drafts customer messages, training notes, and marketing copy, then routes them for review before external use. |
| Unsupported-Answer Guardrail | A confident guess is worse than no answer. Retrieval thresholds, source requirements, and uncertainty rules force the assistant to identify missing support. |
| Claude Product Fit Matrix | Teams often apply one Claude product to every task. The included audit maps chat, coding, API, browser, and spreadsheet work to the most maintainable execution path. |
Sync Internal Wikis With AI Knowledge Assistant Connectors
The ingestion pipeline can sync internal wikis with AI knowledge assistant indexes through MCP or a read-only adapter. Each source receives an owner, location, department, effective date, and access label. Deleted or superseded pages are removed from retrieval during the next scheduled refresh.
The default source refresh runs every 15 minutes, with a Refresh Sources control for immediate updates. Content fetched from connected systems is treated as untrusted tool data so instructions hidden inside a document cannot override the agent policy.
Technology Choices
| Layer | Choice and reason |
|---|---|
| Agent runtime | Claude Messages API handles long operational context, while tool schemas restrict which actions the model may request. |
| Retrieval | Hybrid semantic and keyword search supports exact procedure names, abbreviations, and natural staff questions. |
| Connectors | MCP keeps Workspace, CRM, phone, and document adapters separate from the agent prompt and reusable across modes. |
| Orchestration | Python services validate inputs, enforce permissions, schedule sync jobs, apply retries, and write structured audit events. |
| Interfaces | A browser dashboard supports staff use; a CLI supports maintenance, connector diagnostics, and controlled re-indexing. |
| Storage | A retrieval index stores document chunks, while a separate audit store records source IDs, tool requests, approvals, and response status. |
Validation and Operating Targets
The release gate includes 50 representative clinic questions covering scheduling, staff training, post-visit follow-up, location procedures, and missing-information cases. At least 90% of answerable questions must cite the approved source selected by the evaluator. Unsupported questions must return an escalation response rather than a fabricated procedure.
The reference deployment targets a p95 response time below 8 seconds for indexed questions. Sync failures retry with backoff, appear on the dashboard, and remain visible until acknowledged.
The tool is designed for operational knowledge and drafting. It does not diagnose conditions, prescribe treatment, or replace clinical record systems.
Project Directory
knowledge-base-ai-assistant/
├── app/
│ ├── main.py
│ ├── config.py
│ ├── web/
│ │ ├── routes.py
│ │ ├── auth.py
│ │ └── templates/
│ │ ├── ask.html
│ │ ├── review_queue.html
│ │ └── sources.html
│ ├── agents/
│ │ ├── knowledge_agent.py
│ │ ├── followup_agent.py
│ │ ├── training_agent.py
│ │ └── policies.py
│ ├── retrieval/
│ │ ├── indexer.py
│ │ ├── hybrid_search.py
│ │ ├── chunking.py
│ │ └── citations.py
│ ├── connectors/
│ │ ├── mcp_server.py
│ │ ├── google_drive.py
│ │ ├── google_sheets.py
│ │ ├── webhook_crm.py
│ │ └── phone_adapter.py
│ ├── actions/
│ │ ├── draft_followup.py
│ │ ├── manager_escalation.py
│ │ └── source_refresh.py
│ └── audit/
│ ├── events.py
│ └── redaction.py
├── prompts/
│ ├── system.md
│ ├── employee_mode.md
│ └── manager_mode.md
├── evals/
│ ├── clinic_questions.csv
│ ├── graders.py
│ └── expected_sources.json
├── scripts/
│ ├── sync_sources.py
│ ├── test_connectors.py
│ └── export_audit.py
├── docs/
│ ├── operator-guide.md
│ ├── connector-setup.md
│ └── maintenance-handoff.md
├── tests/
│ ├── test_permissions.py
│ ├── test_retrieval.py
│ ├── test_tool_calls.py
│ └── test_no_answer.py
├── .env.example
├── requirements.txt
└── README.md
Use Cases
- Reduce repeated staff questions: Front-desk and recovery teams receive the same source-linked procedure regardless of location or shift.
- Accelerate onboarding: New staff ask about room preparation, scheduling rules, follow-up steps, and training material without searching multiple folders.
- Prepare manager-reviewed follow-ups: The assistant drafts post-visit messages from approved templates and routes them for review before sending.
- Keep operating guidance current: Managers update one approved document, trigger a refresh, and make the revised procedure available across locations.
Teams extending the project can use CogworkLabs for custom AI assistants for internal knowledge retrieval, connector additions, deployment, and ongoing maintenance.
How to Retrieve Clinic Knowledge Using Knowledge Base AI Assistant
Download & Set Up the Project
Download, set up, and install Knowledge Base AI Assistant to get the project running. If you hit any difficulty, contact us here.
Open the Staff Console
Sign in to the dashboard, choose Employee or Manager mode, and confirm the active clinic location and permitted knowledge collections.
Enter the Request
Type the question, choose Ask, Draft Follow-Up, or Training Summary, then apply optional department, location, date, and source filters.
Run and Review
Select Run Assistant to receive the answer, cited files, confidence status, and any manager-review task created by the workflow.
FAQs
How does the assistant prevent unsupported answers?
It requires retrieved evidence before presenting operational guidance. When no approved source clears the retrieval threshold, the assistant states that the answer is unsupported, records the query, and routes it to the manager queue.
Can it connect to Google Workspace and internal systems?
Yes. The included Drive and Sheets connectors handle approved documents and structured operating data, while MCP or webhook adapters connect compatible CRM, phone, wiki, and internal tools. Each connector can be limited to read-only access or a narrow set of approved actions.
How is an AI knowledge assistant vs chatbot setup different?
A chatbot primarily manages conversation, while a knowledge assistant is built around governed retrieval, source permissions, citations, and no-answer behavior. This tool can converse naturally, but its main job is to find trusted internal knowledge and perform explicitly approved actions.
