# Knowledge Base Open your agent in the console and select **Knowledge Base**. This panel has two tabs: - **Documents** - **Jobs** Related pages: - [Instructions](instructions.md) - [Functions](functions.md) ## What the Knowledge Base does Leilani's Knowledge Base gives an agent a searchable store of reference material it can use when callers ask factual questions or when the agent needs background information before answering. At a high level, the Knowledge Base does three things: - stores uploaded documents for an agent - lets jobs crawl URI-based content and ingest it as documents - exposes both sources through the built-in `knowledge_base_search` tool during calls Knowledge Base content is separate from custom [Functions](functions.md). Functions perform actions. The Knowledge Base supplies retrieval context. ## How Knowledge Base search behaves During a call, Leilani can call `knowledge_base_search` whenever it needs factual grounding or supporting context. In practice: - the caller's question is embedded and matched against stored document chunks - search returns up to three of the closest matches - each match includes the source document name, media type, matching content chunk, and score - uploaded documents and job-created documents are searched together This means jobs do not create a separate index. They create documents that join the same Knowledge Base. ## Documents tab The Documents tab lists all documents for the agent, including documents created manually and documents created by Jobs. The Documents table has these columns: - `Name` - `Media Type` - `Size (MiB)` - `Checksum` - `Created At` Available actions: - `Create` (upload a file) - `Delete` ## Uploading a document Select **Create** to choose a file from your computer. Current document upload behavior: - the console reads the selected file as text and sends that text to the API - the document name is stored in lowercase - the uploaded text is split into overlapping chunks and embedded for search - documents are unique by name within an agent Text-based files work best, such as: - `.md` - `.txt` - `.csv` - `.json` - `.html` Binary formats such as PDF, DOCX, or images are not parsed by this upload path. For website content, use a Job instead. ### Document best practices - keep each document focused on one topic, policy, or FAQ area - use stable, descriptive filenames - prefer cleaned reference text over raw exports - split very large reference sets into smaller files when possible - avoid uploading the same content under multiple names unless you want duplicate retrieval candidates ## Jobs tab The Jobs table has these columns: - `ID` - `URI` - `Created At` - `Updated At` Available actions: - `Create` - `Run` - `Edit` - `Delete` ## Create or edit a Job A Job stores a fully qualified URI that Leilani can process in the background. Current job requirements and behavior: - `URI` is required - the URI must be fully qualified, such as `https://example.com/docs` - `HTTP` and `HTTPS` URIs are currently supported - each agent can only have one Job per URI ## How Jobs run When a Job runs, Leilani: 1. starts at the configured URI 2. crawls reachable pages from that starting point 3. respects `robots.txt` 4. converts crawled HTML into markdown 5. stores each crawled page as a document in the Knowledge Base Job-created documents use: - the final page URL as the document name - `text/markdown` as the media type Before a Job writes new documents, Leilani deletes the documents previously created by that same Job. A rerun is therefore a replace operation, not a merge. You can run a Job manually from the console with **Run**. Leilani also reruns all Knowledge Base Jobs automatically every 24 hours. ### Relationship between Jobs and Documents Jobs do not bypass Documents. They create Documents. That means: - job-created pages appear in the Documents tab - deleting a Job also deletes the documents created by that Job - deleting a job-created document does not delete the Job itself - the next Job run can recreate any deleted job-owned documents ### Job best practices - use Jobs for public or web-hosted reference content that changes over time - start from a narrow documentation section rather than a broad homepage - prefer stable help-center or docs URLs over marketing pages - expect best results from pages that render as normal HTML without authentication - use uploaded Documents for private internal material that is not available by URI ## Choosing between Documents and Jobs Use **Documents** when the source is curated, internal, or already available as clean text. Typical examples: - internal SOPs - pricing notes - policy summaries - office-specific answers - agent scripts or escalation references Use **Jobs** when the source lives on a website and should be refreshed automatically. Typical examples: - public help-center articles - product documentation - knowledge base sections on your own site - reference pages that change over time Many agents should use both: - Documents for private or curated material - Jobs for public or frequently updated web content ## Instruction guidance Knowledge Base quality depends on both content and prompt design. In your [Instructions](instructions.md), tell the agent when it should: - use the Knowledge Base before answering factual questions - summarize retrieved content conservatively - admit uncertainty when results are weak or missing - fall back to web search or another workflow only when appropriate