Expert Categories & Models¶

Detailed description of all 12 expert categories with tool injection and model recommendations.

general — General Knowledge¶

Tier: T2 Use: General knowledge questions, definitions, explanations, summaries

Tool injection: none

Typical models: - T2: gemma3:27b, qwen3.5:35b, llama3.3:70b

System prompt focus: Balanced, informative answers; no subject-area bias.

math — Mathematics¶

Tier: T1 → T2 (at CONFIDENCE < 0.65) Use: Calculations, algebraic equations, statistics, geometry

Tool injection: - calculate – safe arithmetic evaluation - solve_equation – SymPy-based equation solver - prime_factorize – prime factorization - statistics_calc – mean, median, standard deviation

Typical models: - T1: phi4:14b - T2: qwq:32b (reasoning specialist)

Note: MCP tools are injected for exact calculations before the expert responds.

technical_support — IT & DevOps¶

Tier: T1 → T2 Use: System administration, Docker, Kubernetes, Linux, networking, DevOps

Tool injection: - subnet_calc – CIDR/netmask analysis - regex_extract – pattern matching in log files

Typical models: - T1: deepseek-coder-v2:16b - T2: devstral:24b

code_reviewer — Code Analysis¶

Tier: T2 Use: Code review, security analysis, refactoring recommendations, bug hunting

Tool injection: - json_query – structural analysis of configuration files - regex_extract – pattern matching in code

Typical models: - T2: devstral:24b, qwen3-coder:30b

Note: OWASP Top 10 as mandatory checklist in the system prompt.

creative_writer — Creative Writing¶

Tier: T2 Use: Text creation, marketing copy, storytelling, blog articles, emails

Tool injection: none

Typical models: - T2: gemma3:27b, qwen3.5:35b

medical_consult — Medicine¶

Tier: T1 → T2 Use: Medical information, symptom explanations, medication info

Tool injection: none

Typical models: - T1: phi4:14b - T2: gemma3:27b

Note: Critic node checks medical claims for accuracy; mandatory note recommending professional medical consultation.

Disclaimer

Medical information does not replace professional medical advice.

legal_advisor — Law¶

Tier: T2 Use: German law, BGB, StGB, HGB, contract law, criminal law

Tool injection: - legal_search_laws – search for laws - legal_get_paragraph – retrieve paragraph text (exact text) - legal_fulltext_search – full-text search in statutory texts

Typical models: - T2: magistral:24b (law specialist), command-r:35b

Note: Critic node checks for correct paragraph citations; MCP tools provide exact statutory texts.

Disclaimer

Legal information does not replace professional legal advice.

translation — Translation¶

Tier: T2 Use: Professional translations DE↔EN↔FR↔ES↔IT

Tool injection: none

Typical models: - T2: translategemma:27b, qwen3.5:35b

data_analyst — Data Analysis¶

Tier: T1 Use: Statistics, pandas code, SQL queries, data visualization

Tool injection: - statistics_calc – statistical metrics - json_query – dataset analysis

Typical models: - T1: phi4:14b

Note: Fast-path preferred, as requests are often clearly structured.

science — Natural Science¶

Tier: T2 Use: Chemistry, biology, physics, astronomy, climate research

Tool injection: none

Typical models: - T2: gemma3:27b

reasoning — Logic & Analysis¶

Tier: T1 → T2 Use: Complex logic problems, strategic analysis, argumentation structure

Tool injection: none

Typical models: - T1: phi4:14b - T2: deepseek-r1:32b (reasoning chain specialist)

Note: Thinking node is always activated for complex queries.

vision — Image Analysis¶

Tier: T2 Use: Image analysis, screenshot evaluation, document recognition, OCR support

Tool injection: none

Typical models: - T2: Multimodal models (e.g. gemma3:27b, llava:34b)

Note: Base64-encoded images are embedded directly in the request.

# Image request example
import base64

with open("screenshot.png", "rb") as f:
    b64 = base64.b64encode(f.read()).decode()

payload = {
    "model": "moe-orchestrator",
    "messages": [{
        "role": "user",
        "content": [
            {"type": "text", "text": "Describe this image"},
            {"type": "image_url", "image_url": {"url": f"data:image/png;base64,{b64}"}}
        ]
    }]
}