AI Agent Overview

The Vartovii AI Agent ("Vartovii" - Guardian in Ukrainian) is an intelligent assistant powered by Google Gemini 3 Flash preview (with Gemini 2.5 fallback) and Function Calling capabilities.

🤖 What is the AI Agent?

The AI Agent is a conversational interface that can:

• Search companies in our database
• Retrieve Trust Scores with detailed breakdowns
• Compare employers side-by-side
• Start analysis for new companies (Magic Search)
• Show reviews filtered by sentiment
• Answer questions about the platform

🧠 Technology Stack

Component	Technology	Purpose
LLM	Gemini 3 Flash preview (+ 2.5 fallback)	Natural language understanding
Function Calling	Vertex AI	Tool execution
RAG	Vertex AI RAG Engine	Documentation search
Grounding	Google Search	Real-time web information

🔧 How It Works

User Message → Gemini 3 preview → Function Detection → Tool Execution → Response
                    ↓
              Keywords detected?
                ↓         ↓
            Yes: Tool Call    No: Direct Answer
                ↓
          Execute Tool (API call)
                ↓
          Format Response

Example Flow

1. User says: "Find BMW company"
2. Agent detects: Keyword "find" + "company" = search_company tool
3. Agent executes: search_company(company_name="BMW")
4. API returns: Company data with Trust Score
5. Agent responds: Formatted message with company info

🛠️ Available Tools

The agent has access to 12 specialized tools:

The public tool set and response fields stay the same even as the internal chat runtime evolves. Current tool routing is handled through a dedicated internal registry layer so the public contract remains stable. The backend runtime also reuses a shared execution pool for model calls; this is an internal performance and stability improvement and does not change the public AI chat API contract. Tool, grounding, and plain-response model execution now also flow through a dedicated internal runtime adapter layer, again without changing the public AI chat contract. Schema-backed AI tasks also emit internal latency and parse-failure diagnostics, which improves reliability monitoring without changing the public response format. Company-oriented employer tools now also route through a dedicated internal adapter and shared repository layer, which keeps the public responses stable while reducing backend duplication. Crypto project lookup and cached-detail tools now use the same internal adapter pattern, again without changing the public AI chat contract. On-chain forensic tools now follow the same internal adapter approach for wallet, transaction, token-holder, and contract responses. Analysis-triggering flows for new company or project requests now follow the same internal adapter pattern as well. Tool-enabled response assembly now also follows a dedicated internal orchestration adapter, which improves runtime consistency without changing the public AI chat contract. Top-level route selection for tools, RAG, grounding, and standard chat now also follows a dedicated internal routing adapter, again without changing the public AI chat contract. Shared prompt and context composition across standard, grounded, retrieval-backed, and tool-enabled chat paths now also follow a dedicated internal prompt adapter, which improves runtime consistency without changing the public AI chat contract. Core route selection is also regression-tested internally across employer, crypto, forensic, grounded, retrieval-backed, and direct-answer chat paths. Representative report, crypto analysis, and forensic outputs are also guarded internally by canonical gold fixtures so the public AI chat and report contracts stay stable as the runtime evolves. Internal CI now also publishes a dedicated regression summary for those routing and gold-fixture suites. Internal CI also publishes focused coverage summaries for the canonical AI runtime core modules.

Corporate Intelligence (6 tools)

Tool	Purpose	Example Trigger
search_company	Find company in database	"Find BMW"
get_trust_score	Get detailed Trust Score	"What's BMW's Trust Score?"
list_companies	Show top employers	"Show top 10 companies"
compare_companies	Compare two employers	"Compare BMW and Audi"
run_analysis	Start new company analysis	"Analyze Spotify"
get_company_reviews	Show sample reviews	"Show BMW reviews"

Crypto Intelligence (2 tools)

Tool	Purpose	Example Trigger
search_crypto_projects	Find crypto project (CoinGecko)	"Find Solana"
get_crypto_trust_score	Get Crypto Trust Score	"Trust Score for ETH"

Forensic Agent (4 tools)

Tool	Purpose	Example Trigger
check_wallet	Check ETH balance for wallet	"Balance of 0xd8dA..."
get_transaction_history	Get wallet transaction history	"Transactions for 0xd8dA..."
get_token_holders	Analyze token holder distribution	"Top holders of USDT"
get_contract_info	Verify smart contract details	"Is 0xdAC17F... ERC20?"

→ See Tools Reference for complete documentation.

🔍 Auto-Detection

The agent automatically identifies when to use tools based on keywords:

English Keywords

• Corporate: company, employer, find, search, reviews
• Crypto: crypto, token, coin, price, market cap, project
• Forensic/On-Chain: wallet, address, balance, transaction, holder, contract, 0x
• Analysis: analyze, scrape, trust score, rating
• Comparison: compare, versus, vs, battle

Ukrainian Keywords

• Корпоративні: компанія, роботодавець, відгуки, рейтинг
• Крипто: крипта, токен, монета, ціна, проект
• Форензік: гаманець, адреса, баланс, транзакції, контракт
• Аналіз: проаналізуй, збери дані, траст скор
• Порівняння: порівняй, проти, vs

💬 Chat Widget

The AI Agent is accessible via a floating chat button (💬) in the bottom-right corner of the dashboard.

Features

• English UI by default - Chat interface fully localized to English
• Multi-language responses: English, Ukrainian, German (auto-detected)
• Conversation history: Context maintained across messages
• Markdown rendering: Tables, code blocks, lists
• Tool status visualization: Pizza Tracker shows active tools

📊 Pizza Tracker

When the agent uses tools, the Pizza Tracker shows real-time progress:

🔍 search_company → Searching for "BMW"...
✅ Found: BMW (Trust Score: 71)

The tracker displays:

• Tool name being executed
• Parameters passed
• Execution status (pending/complete)
• Result summary

🔐 Security

• API Key Required: Agent tool endpoints require X-API-Key header
• Rate Limited: 5 requests/minute for analysis tools
• Audit Logged: All tool calls are logged for security

🚀 Getting Started

Via Chat Widget

1. Click the 💬 button in the dashboard
2. Type your question in natural language
3. Agent will detect intent and execute appropriate tools

Via API

curl -X POST https://sentryanalytic.com/api/ai/chat \
  -H "Content-Type: application/json" \
  -d '{
    "message": "Find BMW company",
    "language": "auto"
  }'

📝 Response Format

{
  "response": "BMW is a German automotive company with a Trust Score of 71/100...",
  "status": "ok",
  "grounded": false,
  "rag_used": false,
  "tools_used": true,
  "tool_calls": [
    { "name": "search_company", "args": { "company_name": "BMW" } }
  ]
}

---

See Tools Reference for detailed documentation of each tool.