Switching LLM providers¶

Every architecture in this library is provider-agnostic. The same Reflection class works against Nebius Llama, OpenAI GPT-4o, Anthropic Claude, Groq, Ollama, Together, Fireworks, Mistral, or Google Gemini — you change one config setting.

The `get_llm()` factory¶

from agentic_architectures import get_llm

# Use whatever's in .env / settings
llm = get_llm()

# Override per-call
llm = get_llm(provider="openai", model="gpt-4o-mini", temperature=0.4)
llm = get_llm(provider="anthropic", model="claude-haiku-4-5-20251001")
llm = get_llm(provider="ollama", model="llama3.3:70b")

Pass the returned LLM into any architecture:

from agentic_architectures.architectures import Reflection
arch = Reflection(llm=llm)

Provider compatibility matrix¶

Provider	Extra to install	Env var	Notes
Nebius (default)	`[nebius]`	`NEBIUS_API_KEY`	Llama-3.3-70B + Qwen3-Thinking. Cheapest for the included demos.
OpenAI	`[openai]`	`OPENAI_API_KEY`	All architectures work. Highest quality for reasoning architectures.
Anthropic	`[anthropic]`	`ANTHROPIC_API_KEY`	Strong on long context. Required for true Computer-Use (nb 34) production.
Groq	`[groq]`	`GROQ_API_KEY`	Fast inference; good for high-volume Self-Consistency (nb 21).
Ollama	`[ollama]`	—	Fully local; no API key. Tool calling depends on model — Llama-3.3 instruct works.
Together	`[together]`	`TOGETHER_API_KEY`	Wide model catalogue.
Fireworks	`[fireworks]`	`FIREWORKS_API_KEY`	Function-calling first-class.
Mistral	`[mistral]`	`MISTRAL_API_KEY`	EU-hosted option.
Google	`[google]`	`GOOGLE_API_KEY`	Gemini-2.x via Generative AI API.

Environment configuration¶

Settings load from .env at the repo root. Minimum:

# .env
LLM_PROVIDER=nebius
LLM_MODEL=meta-llama/Llama-3.3-70B-Instruct
LLM_TEMPERATURE=0.2

NEBIUS_API_KEY=sk-...
TAVILY_API_KEY=tvly-...
LANGSMITH_API_KEY=ls-...           # optional, for tracing
LANGSMITH_PROJECT=all-agentic-architectures

To switch globally, change LLM_PROVIDER + the corresponding key. No code changes needed.

When you'd override per-architecture¶

Reasoning architectures (nb 19, 20, 21, 22) default to Qwen3-Thinking for higher quality on multi-step structured reasoning. If you want them on a cheaper model:

from agentic_architectures.architectures import LATS
arch = LATS(llm=get_llm(model="meta-llama/Llama-3.3-70B-Instruct"))

Or the opposite — force a single architecture onto a stronger model:

from agentic_architectures.architectures import ChainOfVerification
arch = ChainOfVerification(llm=get_llm(provider="openai", model="gpt-4o"))

Capability matrix¶

Feature	Required for	Providers known good
`with_structured_output(Schema)`	nb 04, 06, 09, 10, 13, 15, 17-35	All listed above
`bind_tools([...])`	nb 02, 03, 05, 11	All except some Ollama tool-uncapable models
`bind(temperature=...)` per-call override	nb 21 Self-Consistency	All

If an architecture fails on a provider, the failure is usually a missing capability — see the notebook's §10 "Try other providers" cell for graceful-skip patterns.