> ## Documentation Index > Fetch the complete documentation index at: https://docs.vendschat.com/llms.txt > Use this file to discover all available pages before exploring further. # Knowledge sources & testing AI Agents > Give agents accurate answers from files and websites, then validate in the Test AI Agent sandbox before customers see replies. Give agents accurate answers from **files** and **websites**, then validate in the **Test AI Agent** sandbox before customers see replies. **Where:** **Train AI Agent** / **Edit AI Agent** form → **Knowledge Sources** section + **Test AI Agent** panel *** ## Quick reference | Task | Where | | ----------------- | ---------------------------------------------------- | | Add file | **Add knowledge source** → **Add file** | | Add website | **Add knowledge source** → **Add website** | | Remove source | Trash icon on row | | Manual resync | Refresh icon (edit mode, after save) | | Test replies | **Test AI Agent** panel — type + send | | View sources used | Test replies include retrieval **sources** (backend) | *** ## Why knowledge matters Banner on form and detail: > *“Want your AI Agent to respond more accurately? Add accurate knowledge sources…”* The model uses **retrieved snippets** from your sources for factual answers — not just instructions. Update knowledge after every **pricing**, **policy**, or **product** change. Vendschat Train AI Agent form showing Response Delay and Knowledge Sources with Add knowledge source, plus Create agent. *** ## Add knowledge source drawer **Button:** **Add knowledge source** → drawer **“Add AI Knowledge Source”** ### Shared fields | Field | Notes | | -------- | ------------------------------------------ | | **Name** | Label in list (e.g. `Returns policy 2024`) | ### Option A: File | Rule | Value | | ------------------- | --------------------------------------------------------------- | | Max files per batch | **5** | | Max per file | **1 MB** | | Max combined batch | **1 MB** total | | Allowed types | PDF, TXT, MD, CSV, DOCX, PPTX, PPSX, JPEG, PNG, BMP, WEBP, TIFF | Upload flow uploads to workspace storage → attaches `file_id` on save. ### Option B: Website | Field | Purpose | | ------------------- | -------------------------------- | | **Website URL** | Starting URL to crawl | | **Resync schedule** | Never · Daily · Weekly · Monthly | | **Resync time** | Time of day (when not Never) | | **Day of week** | Weekly schedule | | **Day of month** | Monthly schedule (days 1–28) | #### Advanced crawl settings | Setting | Default | Purpose | | ---------------------- | ------- | --------------------------- | | **Load sitemap** | On | Discover pages from sitemap | | **Max crawling depth** | 3 | Link levels from start URL | | **Additional URLs** | Up to 5 | Extra seeds | | **Include globs** | — | Only crawl matching paths | | **Exclude globs** | — | Skip matching paths | Path filter docs link in UI points to crawl path filter documentation. *** ## Source list (after add) Each row shows: | Type | Display | | ----------- | --------------------------------------- | | **File** | Filename, size | | **Website** | URL + resync schedule label | | **Status** | `ready` · `syncing` · `error` | | **Resync** | Refresh button when editing saved agent | | **Remove** | Delete from agent config | **Added sources (N)** counter on form. On save (create or edit), sources sync to backend indexing pipeline. *** ## Rate limits (fair use) Rolling **24-hour** window per workspace: | Limit | Default | Error message theme | | ------------------------------------------ | ------- | --------------------------------- | | Knowledge source **changes** per workspace | 50/day | Wait 24h or reduce changes | | **Crawl/resync** per source | 50/day | Wait before resyncing same source | If customer hits limit, wait or contact support for enterprise limits. *** ## How retrieval works (conceptual) ```text theme={null} Customer question → hybrid search over indexed chunks for THIS agent → top snippets injected into prompt → model answers with [Source: name] style citations in prompt guidance ``` Conversation history (recent messages) is included separately from knowledge files. **Test mode** returns `sources` array with heading, content, score for debugging. *** ## Test AI Agent panel Available on: * **Create / Edit** form (side panel / mobile drawer) * **Agent detail** page (**Test AI Agent** button) Vendschat Sales Agent detail on the left and Test AI Agent sandbox on the right with a greeting and a Type your message field. ### UI behavior | Element | Purpose | | ------------------- | ------------------------------------------------------- | | Chat bubbles | User (outgoing style) vs assistant (incoming) | | Welcome line | From `welcome_message` if set, else agent name greeting | | Input + send | Type customer questions | | **Reset** | Clear test conversation | | Mic / sticker icons | Visual parity with Chat (test only) | ### What to test | Category | Example messages | | ------------------- | ------------------------------------------------------- | | FAQ | *“What are your hours?”* | | Policy edge | *“Can I return worn shoes?”* | | Escalation | *“I want to speak to a manager”* | | Language | Customer language you support | | Hallucination check | Obscure SKU not in knowledge — should admit uncertainty | | Actions | Phrasing that should trigger assign/label rules | **Test does not** send WhatsApp/Messenger messages — sandbox only. ### After failed test Error toast; last user message restored to input for retry. *** ## Knowledge status badges | Status | Meaning | | ----------- | ------------------------ | | **ready** | Indexed and searchable | | **syncing** | Crawl/upload in progress | | **error** | Fix URL/file; try resync | Manual **Resync** queues a new ingest (subject to rate limits). *** ## Example: Harbor Legal knowledge stack | Source | Type | Resync | | --------------------- | -------- | -------------------- | | `Intake FAQ` | PDF file | — | | `harborlegal.com/faq` | Website | Weekly, Monday 00:00 | | `Fee schedule` | DOCX | — | Test: *“Do you handle immigration?”* → should cite FAQ, not invent practice areas. *** ## Best practices 1. **One topic per file** when possible — easier to update 2. **Exclude** blog/noise URLs with exclude globs 3. **Weekly resync** for pricing pages; **monthly** for static policies 4. **Re-test** after every knowledge edit 5. **Don’t upload** huge dumps — stay within 1 MB batches; split logically 6. **Instructions** repeat what knowledge cannot enforce (tone, legal boundaries) *** ## Troubleshooting | Issue | Fix | | ---------------------- | ----------------------------------------------------- | | Upload rejected | Check 1 MB limit and file type list | | Website stuck syncing | Valid public URL; check robots; retry resync tomorrow | | Bot invents facts | Add/fix knowledge; tighten instructions | | Test works, live fails | Thread assigned to agent? Agent **Active**? | | Rate limit 429 | Wait 24 hours | | No resync button | Save agent first (edit mode with document id) | *** ## Related docs * [Creating & training agents](/ai-agents/creating-and-training-agents) * [AI Agents overview](/ai-agents/ai-agents-overview) * [Assigning agents in Chat](/ai-agents/assigning-agents-in-chat) * [AI Agents FAQ](/ai-agents/ai-agents-faq)