farabi-0.6B-agent-rag

Compact trilingual (Kazakh / Russian / English) language model for RAG and tool-calling agents.

Capabilities

• RAG / grounded QA — answers from provided context chunks; cites sources; abstains when evidence is insufficient
• Tool calling — Hermes-format; clarifies missing required arguments before calling; recovers gracefully from tool errors
• Claim verification — SUPPORTED / REFUTED / NOT_ENOUGH_INFO against provided passages
• Rule & numeric compute — business-day arithmetic, fee/eligibility calculations from provided rules
• Multilingual — Kazakh, Russian, English; handles cross-lingual RAG queries

Serving with vLLM

vllm serve nur-dev/farabi-0.6B-agent-rag \
  --enable-auto-tool-choice \
  --tool-call-parser hermes \
  --chat-template chat_template.jinja \
  --max-model-len 8192

OpenAI Agents SDK

import openai

client = openai.AsyncOpenAI(
    base_url="http://localhost:8000/v1",
    api_key="x",
)

response = await client.chat.completions.create(
    model="farabi-0.6B-agent-rag",
    messages=[{"role": "user", "content": "..."}],
    tools=[...],
    tool_choice="auto",
)