DLE/docs.en/back-docs/setup-ai-assistant.md

English | Русский

AI Assistant Setup with Vector Search

Full guide to launching the intelligent assistant

This document describes the step-by-step setup of the AI assistant for business tasks using spreadsheets and vector search.

---

What you will have after setup

✅ Working AI assistant with local model (Ollama)
✅ Knowledge base for customer answers (FAQ)
✅ Supplier and procurement automation
✅ Staff training system
✅ Vector search over your data
✅ Significant time and cost savings

💡 Economics: See DLE AI Agents for architecture, examples, and savings.

---

Time required

• Quick setup: 20–30 minutes (basic FAQ)
• Full setup: 1–2 hours (all features)

---

Step 1: Install and run Ollama

1. Settings → Integrations → Ollama → Details
2. Check status: “Ollama is running” or “Ollama API not responding”
3. If not running: docker-compose up -d ollama or ollama serve
4. Install model: e.g. qwen2.5:7b (recommended), llama2:7b, mistral:7b
5. Install embedding model: mxbai-embed-large:latest (recommended) or nomic-embed-text:latest

⚠️ Embedding model is required for RAG (vector search).

---

Step 2: Create knowledge base (spreadsheets)

2.1 FAQ table

1. Tables → + Create table
2. Name: e.g. “FAQ – Frequently asked questions”, description for AI
3. Add columns:
   • Question — type Text, purpose: Question for AI (required for RAG)
   • Answer — type Text, purpose: Answer for AI
   • Product (optional) — Multiselect, purpose: Product filter
   • Tags (optional) — Multiselect, purpose: User tags
   • Priority (optional) — Number, purpose: Priority

2.2 Fill with sample Q&A

Add rows: e.g. “How to pay?” / “We accept card, PayPal, bank transfer…”; “Delivery time?” / “3–5 business days…”; “Return policy?” / “Within 14 days…”. Minimum ~20–30 questions recommended.

2.3 Enable as AI source

In table settings enable “Use as source for AI” and save. Table is then indexed for vector search.

---

Step 3: AI provider (Ollama) settings

1. Settings → Integrations → Ollama
2. Base URL: Docker http://ollama:11434, local http://localhost:11434
3. LLM model: e.g. qwen2.5:7b
4. Embedding model: mxbai-embed-large:latest
   Save.

---

Step 4: AI Assistant settings

1. Settings → Integrations → AI Assistant → Details
2. System prompt — e.g. “You are a professional support assistant. Answer from the knowledge base. If not found, suggest contacting an operator. Always end with ‘How else can I help?’”
3. Models: select same LLM and embedding as above
4. Selected RAG tables: choose your FAQ table
5. Rules (JSON): e.g. searchRagFirst: true, generateIfNoRag: true, temperature: 0.7, maxTokens: 500
6. RAG search: e.g. Hybrid, max results 5, relevance threshold 0.1; optional keyword extraction, fuzzy search, stemming
   Save.

---

Step 5: Test

1. RAG tester (on assistant settings page): choose table, ask e.g. “How to pay?” → check answer and score (good: about -300 to 0).
2. Web chat: open main page, ask e.g. “What is the delivery cost?” — answer should come from your FAQ.
3. Try questions inside and outside the knowledge base; test with typos (fuzzy search).

---

Step 6 (optional): Extra tables and channels

• Suppliers table: columns for company, category, contact, email, phone, prices, payment terms, delivery, rating. Enable as AI source; add prompt instructions for “TOP-3 suppliers” style answers.
• Staff knowledge base: questions/answers by category (Sales, HR, IT). Same RAG setup.
• Telegram: create bot via @BotFather, add token and username in Settings → Integrations → Telegram; link to AI assistant.
• Email: IMAP/SMTP in Settings; for Gmail use app password. Link to AI assistant.

---

Step 7: Monitoring and tuning

• Status: Settings → AI Assistant → Monitoring: Backend, Postgres, Ollama, Vector Search should be green.
• RAG quality: Score -300…0 = good; >300 = not found. Improve by adding variants of questions and adjusting relevance threshold.
• Speed: Smaller model or fewer RAG results if responses are slow.

---

Troubleshooting

• Ollama not responding: docker-compose restart ollama, check logs.
• Wrong answers: Check RAG score; add more questions; lower relevance threshold; ensure column purposes “Question for AI” and “Answer for AI”.
• Vector search error: Install embedding model; on table page use “Rebuild index”; ensure table is enabled as AI source.
• Wrong language: Add “Always answer in English” (or desired language) to system prompt; choose suitable model (e.g. qwen2.5:7b for multilingual).

---

Technical reference (developers)

• DB: ai_providers_settings, ai_assistant_settings, ai_assistant_rules (encrypted fields, RAG tables, rules JSON).
• API: GET/PUT settings per provider and assistant; rules CRUD; Ollama status, models, install.
• Flow: Message → UnifiedMessageProcessor → language check → dedup → load settings and rules → RAG search → generate LLM response → return. Security: AES-256 for sensitive fields; admin-only for settings.

---

© 2024-2026 Alexander Viktorovich Tarabanov. All rights reserved.

Version: 1.0.0 | Date: February 28, 2026