189 lines
7.1 KiB
Markdown
189 lines
7.1 KiB
Markdown
**English** | [Русский](https://hb3-accelerator.com/gitea/VC-HB3-Accelerator/Docs/src/branch/main/docs.ru/ai-assistant.md)
|
||
|
||
# DLE AI Agents — Building Specialized Business Agents
|
||
|
||
> **Concept:** one local model — many specialized agents. Each agent is tailored to a specific business process: its own prompt, rules, knowledge base, and interface.
|
||
|
||
## Table of Contents
|
||
|
||
1. [What and Why](#what-and-why)
|
||
2. [Architecture](#architecture)
|
||
3. [How to Create an Agent](#how-to-create-an-agent)
|
||
4. [Agent Examples](#agent-examples)
|
||
5. [Technology Stack](#technology-stack)
|
||
6. [Advantages Over Cloud Solutions](#advantages-over-cloud-solutions)
|
||
7. [Economic Impact](#economic-impact)
|
||
|
||
---
|
||
|
||
## What and Why
|
||
|
||
DLE provides **tools to create AI agents** — specialized assistants, each responsible for a distinct business process.
|
||
|
||
This is not one generic chatbot. It is a **builder** where you:
|
||
|
||
- Create an agent for a specific task (support, content, procurement, analytics)
|
||
- Set its role via system prompt
|
||
- Attach a knowledge base (RAG tables) with relevant data
|
||
- Configure behavior rules (strict, creative, hybrid)
|
||
- Bind to channels (web chat, Telegram, Email)
|
||
- Get an isolated specialist working 24/7
|
||
|
||
All agents use **one local Ollama model** on your server. They differ by system prompts, rules, and connected data. Data never leaves your server.
|
||
|
||
---
|
||
|
||
## Architecture
|
||
|
||
### Principle: One Model — Many Agents
|
||
|
||
One Ollama instance (e.g. qwen2.5:7b) serves multiple agents. Each has its own prompt, rules, RAG tables, channels, and UI. Agents are isolated and do not affect each other.
|
||
|
||
### Request Flow
|
||
|
||
User request → agent determined by channel/route → agent config (prompt, rules, RAG) → query vectorization (Ollama mxbai-embed-large) → RAG search (FAISS) → LLM response with RAG context + system prompt + history → optional cache (TTL 1 h) → response to user.
|
||
|
||
---
|
||
|
||
## How to Create an Agent
|
||
|
||
### Step 1. Basic Info
|
||
|
||
- **Name** — e.g. “Support Agent”, “Content Editor”, “AI Procurement”
|
||
- **Role** — support, content_editor, analyst, purchaser, etc.
|
||
- **Description** — what the agent is for
|
||
|
||
### Step 2. System Prompt
|
||
|
||
Defines identity and behavior. Examples:
|
||
|
||
**Support:** “You are a professional customer support assistant. Answer only from the knowledge base. If no answer — suggest contacting an operator. Do not invent prices, terms, or conditions.”
|
||
|
||
**Content editor:** “You are a professional content marketer and editor. Use company style from the knowledge base. Follow platform guidelines. Use keywords and hashtags from the base.”
|
||
|
||
### Step 3. Rules (JSON)
|
||
|
||
```json
|
||
{
|
||
"searchRagFirst": true,
|
||
"generateIfNoRag": false,
|
||
"checkUserTags": true,
|
||
"temperature": 0.3,
|
||
"maxTokens": 500
|
||
}
|
||
```
|
||
|
||
| Parameter | Effect | Support | Content | Analytics |
|
||
|-----------|--------|----------|---------|-----------|
|
||
| temperature | Creativity (0.0–1.0) | 0.3 | 0.7 | 0.2 |
|
||
| searchRagFirst | Search knowledge base first | true | true | true |
|
||
| generateIfNoRag | Generate if not in base | false | true | false |
|
||
| maxTokens | Max response length | 500 | 2000 | 1000 |
|
||
|
||
### Step 4. Knowledge Base (RAG Tables)
|
||
|
||
Attach spreadsheets the agent will search: Support → FAQ, product docs; Content → platform instructions, style, examples, keywords; Procurement → supplier base, terms, prices. Tables need columns designated as “Question for AI” and “Answer for AI” for vector indexing.
|
||
|
||
### Step 5. Channels and Interface
|
||
|
||
Channels: web chat, Telegram, Email, SMS. Route: e.g. `/content-editor`. Set which roles can access.
|
||
|
||
### Step 6. Activate
|
||
|
||
Enable the agent; it starts handling requests on the selected channels.
|
||
|
||
---
|
||
|
||
## Agent Examples
|
||
|
||
### 1. Customer Support Agent
|
||
|
||
**Task:** answer customer questions 24/7 from the knowledge base. Strict mode (only from base), temperature 0.3, RAG: FAQ, docs. Channels: web chat, Telegram, Email. If no answer → suggest operator.
|
||
|
||
### 2. Content Editor Agent
|
||
|
||
**Task:** create social posts, blog articles, emails in company style. Creative mode, temperature 0.7, RAG: platform instructions, style, examples, keywords, CTAs. Interface: `/content-editor`, Editor role.
|
||
|
||
### 3. AI Procurement Agent
|
||
|
||
**Task:** help choose suppliers and compare terms. Hybrid mode, temperature 0.5, RAG: supplier base, terms and prices. Example: “Who supplies electronics with delivery up to 3 days?” → top 3 from table with filters.
|
||
|
||
### 4. Other Possible Agents
|
||
|
||
Analyst (reports, trends), HR assistant (screening, policies), Translator (glossaries, style), Legal assistant (contracts, norms). Each = new combination of prompt, rules, and tables.
|
||
|
||
---
|
||
|
||
## Technology Stack
|
||
|
||
| Component | Technology | Purpose |
|
||
|-----------|------------|---------|
|
||
| LLM | Ollama (qwen2.5:7b or other) | Generation, dialogue |
|
||
| Embedding | mxbai-embed-large | Text vectorization |
|
||
| Vector DB | FAISS | Semantic search |
|
||
| Main DB | PostgreSQL | Agents, knowledge, history |
|
||
| Cache | Node.js Map + TTL | Fast repeat queries (< 50ms) |
|
||
| Queue | AI Queue | Priority processing |
|
||
| Encryption | AES-256 | Prompts and settings encrypted |
|
||
|
||
### RAG Search Methods
|
||
|
||
Semantic (FAISS), keyword, hybrid (e.g. 70% semantic, 30% keyword). Optional: fuzzy search, stemming, keyword extraction.
|
||
|
||
---
|
||
|
||
## Advantages Over Cloud Solutions
|
||
|
||
| | DLE (local) | ChatGPT API | Claude API |
|
||
|-|-------------|-------------|------------|
|
||
| **Cost** | $0 | ~$0.02/request | ~$0.03/request |
|
||
| **Confidentiality** | 100% on your server | Data at OpenAI | Data at Anthropic |
|
||
| **Speed (cached)** | < 50ms | 500–2000ms | 500–2000ms |
|
||
| **Offline** | Yes | No | No |
|
||
| **Business tuning** | Full: prompts, rules, RAG | Limited | Limited |
|
||
| **API limits** | None | Yes | Yes |
|
||
| **Number of agents** | Unlimited | Separate API per use | Separate API per use |
|
||
|
||
---
|
||
|
||
## Economic Impact
|
||
|
||
### API Cost Savings
|
||
|
||
| Requests/month | ChatGPT API | Claude API | DLE |
|
||
|----------------|-------------|------------|-----|
|
||
| 10,000 | $2,400/year | $3,600/year | $0 |
|
||
| 50,000 | $12,000/year | $18,000/year | $0 |
|
||
| 100,000 | $24,000/year | $36,000/year | $0 |
|
||
|
||
### Process Automation Savings
|
||
|
||
Each agent replaces routine work. Example for a mid-size company: support agent ($57,600), procurement ($64,800), HR ($57,600), content ($86,400), analyst ($144,000), partners ($43,200), training ($30,000), API savings ($24,000–36,000) → **total about $507,600–519,600/year**. DLE license: $1,000 one-time.
|
||
|
||
### 5-Year Comparison with SaaS
|
||
|
||
Typical SaaS stack (CRM, chatbot, email, AI API): ~$39,000 over 5 years. DLE: $1,000 + $0 AI + free updates 5 years → **savings about $38,000**.
|
||
|
||
---
|
||
|
||
## Additional Materials
|
||
|
||
- [Multi-agent architecture](./back-docs/multi-agent-architecture.md) — technical spec
|
||
- [AI assistant setup](./back-docs/setup-ai-assistant.md) — deployment steps
|
||
- [Tables system](./back-docs/tables-system.md) — RAG tables
|
||
- [FAQ](https://github.com/VC-HB3-Accelerator/.github/blob/main/en/FAQ.md)
|
||
|
||
---
|
||
|
||
## Support
|
||
|
||
- **Email:** info@hb3-accelerator.com
|
||
- **Chat:** https://hb3-accelerator.com
|
||
- **Docs:** [FAQ](https://github.com/VC-HB3-Accelerator/.github/blob/main/en/FAQ.md)
|
||
|
||
---
|
||
|
||
**© 2024-2026 Alexander Viktorovich Tarabanov. All rights reserved.**
|
||
|
||
**Last updated:** February 2026
|