Quick Start Guide

Create an AI agent and query it.

Prerequisites

• Twig account at app.twig.so

• Data to index: PDF/Word files, or a documentation website URL

Step 1: Create Your Organization

On first login:

1. Go to app.twig.so

2. Sign in with email/password or SSO (Google, Microsoft)

3. Fill the organization setup form:

   • Organization name (required)

   • Industry (optional dropdown)

   • Invite team members (optional, can skip)

4. Click Complete Setup

Expected result: You see the main dashboard with "Data", "Agents", "Playground" menu items

Step 2: Add a Data Source

Option A: Upload a File

1. Click Data in left navigation

2. Click Add Data Source button (top right)

3. Select File Upload from the modal

4. Choose file: PDF, DOCX, TXT (max 50MB per file)

5. Click Upload

6. Wait for processing (progress bar shows chunking, embedding)

Expected result: Status changes to "Active" with green indicator. Shows chunk count (e.g., "256 chunks indexed").

Processing time: ~1-3 minutes for a 100-page PDF

Option B: Connect a Website

1. Click Data in left navigation

2. Click Add Data Source button

3. Select Website Crawler

4. Enter base URL: https://docs.example.com

5. Set max pages (default: 100, max: 10,000)

6. Click Start Crawl

Expected result: Status "Crawling" → "Processing" → "Active". Shows pages crawled count.

Processing time: ~5-10 minutes for 100 pages

Other Connectors

• Confluence: OAuth flow, select spaces

• Google Drive: OAuth flow, select folders

• Slack: OAuth flow, select channels (last 90 days of messages)

See Data Sources for connector-specific setup

Step 3: Create an Agent

1. Click Agents in left navigation

2. Click Create Agent button

3. Fill the form:

   • Name: e.g., "Customer Support Agent" (required)

   • Data Sources: Check the data source from Step 2 (required)

   • System Prompt (optional): "You are a customer support assistant. Provide concise answers with citations."

   • RAG Strategy: Leave as "Redwood" (default)

   • Model: Leave as "GPT-4" (default)

4. Click Create Agent

Expected result: Agent appears in agents list with status "Active". Agent ID is visible (format: agent_abc123).

Step 4: Test Your Agent

1. Click Playground in left navigation

2. Select your agent from the dropdown

3. Type a question in the input field (must relate to your indexed data)

4. Press Enter or click Send

Expected result:

• Response appears in chat window (typical: 2-5 seconds)

• Citations shown below response with document names and chunk IDs

• Retrieved sources panel (right side) shows the chunks used

How to verify it works:

• Click a citation → opens source document at that location

• Check "Sources Used" panel → shows 5-10 chunks with similarity scores (0.0-1.0)

• Response references content from your data (not generic AI knowledge)

What's Next?

Improve Agent Accuracy

• Add data sources: Agents → [Your Agent] → Data Sources tab → Add more

• Tune system prompt: Agents → [Your Agent] → Configuration → edit System Prompt field

• Change RAG strategy: Agents → [Your Agent] → Configuration → RAG Strategy dropdown

  • Redwood: 1-2s latency (current)

  • Cedar: 2-3s latency, query rewriting

  • Cypress: 3-5s latency, reranking, best accuracy

• Test queries: Run 10-20 test questions in Playground, check citation accuracy

Deploy

• REST API: Settings → API Keys → Generate, use /api/v1/query endpoint

• Embed widget: Deployments → Widget → Copy embed code (iframe)

• Chrome extension: Chrome Web Store → "Twig AI" → Install, paste API key

• Slack integration: Integrations → Slack → Authorize workspace

Monitor

• Inbox: Review → Inbox → see all queries, mark accurate/inaccurate

• Analytics: Dashboard → shows query count, avg latency, accuracy rate

• Evals: Evaluation → Create Test Set → add question/expected answer pairs, run against agent

Troubleshooting

Data Source Stuck at "Processing"

Symptom: Status stays "Processing" for >10 minutes

Diagnostic steps:

1. Data → [Your Data Source] → Logs tab → check for error messages

2. Common errors:

   • "401 Unauthorized" → OAuth token expired, reconnect

   • "Rate limit exceeded" → Wait 1 hour, crawler resumes automatically

   • "Invalid file format" → Convert to PDF/DOCX, re-upload

Fix: Based on error message. If no error after 30 minutes, contact support with data source ID.

Agent Returns "No relevant information found"

Symptom: Agent responds with "I don't have information about that"

Diagnostic steps:

1. Playground → Check "Sources Used" panel → if empty, retrieval failed

2. Possible causes:

   • Query doesn't match indexed content (semantic mismatch)

   • Data source not connected to agent

   • Similarity threshold too high (default: 0.7)

Fix:

• Agents → [Your Agent] → Data Sources tab → verify data source is checked

• Try exact phrase from your document in query

• Lower similarity threshold: Configuration → Advanced → Similarity Threshold → 0.5

Query Takes >10 Seconds

Symptom: Response latency exceeds 10 seconds

Diagnostic steps:

1. Playground → Query → Check response time breakdown (bottom of response)

2. Identify slow step:

   • Retrieval >3s → Too many chunks indexed, contact support

   • LLM >8s → Model overloaded, retry in 1-2 minutes

   • Embedding >1s → Network issue, check connection

Fix: Switch to Redwood strategy (Configuration → RAG Strategy) for 1-2s responses.

When This Doesn't Apply

These steps assume default settings on Free/Pro plan. Enterprise configurations may differ.

---

Next steps: Authentication for API access, or Core Concepts for RAG terminology.