LLM providers¶
Contractex never chooses a provider or a model for you. Every component that calls a model needs one passed in explicitly. The core install includes no vendor SDK: install the extra for the provider you use.
| Provider | Install | Class | Needs |
|---|---|---|---|
| Ollama (runs locally) | pip install "contractex[ollama]" | LocalProvider | Ollama running, the model pulled (ollama pull <model>); OLLAMA_HOST if not http://localhost:11434 |
| OpenAI | pip install "contractex[openai]" | OpenAIProvider | OPENAI_API_KEY |
| Anthropic | pip install "contractex[anthropic]" | AnthropicProvider | ANTHROPIC_API_KEY |
pip install "contractex[google]" | GoogleProvider | GOOGLE_API_KEY | |
| LangChain | pip install "contractex[langchain]" | LangChainProvider | a LangChain chat model instance |
Only LocalProvider is treated as local by the privacy router, so it is the only provider allowed for restricted documents.
Model names¶
These docs do not list cloud model identifiers, because they change. Use the identifier from your vendor's current list: OpenAI, Anthropic, Google, Ollama. The only identifier used in these docs is the Ollama tag llama3.1:8b, in the local examples.
from contractex.llm import AnthropicProvider, LocalProvider, OpenAIProvider
local = LocalProvider(model="llama3.1:8b")
openai = OpenAIProvider(model="your-openai-model-id")
anthropic = AnthropicProvider(model="your-anthropic-model-id")
Anywhere a provider is accepted, you can pass a model name instead. A name beginning gpt- creates an OpenAIProvider, claude- an AnthropicProvider, and anything else a LocalProvider for that Ollama model. A vendor name on its own is rejected rather than mapped to a model the library picks:
from contractex import ContractExtractor
try:
ContractExtractor(llm_provider_name="openai")
except ValueError as exc:
print(exc)
'openai' does not name a model. Pass a full model name such as 'gpt-...' or 'claude-...', or an LLMProvider instance.
Writing a provider¶
Subclass LLMProvider and implement complete, extract_structured, estimate_cost, count_tokens, context_window and model. The privacy guide has a complete example, a provider that prints its prompts, which is also useful for testing.
Known issues¶
- Anthropic: the extra is pinned to
anthropic<1.0. Version 1.0 of the SDK removed thetemperatureparameter thatAnthropicProviderpasses on every call. - Google:
GoogleProvideruses thegoogle-genaiSDK and falls back to the deprecatedgoogle-generativeaiif that is what is installed. Neither path is tested against the live API in this repository. - Cost estimates:
estimate_cost()uses price tables built into each provider class, which date from 2024. Treat them as rough and check your vendor's pricing. - Determinism: model output varies between runs, even at temperature 0. Measure quality with the evaluation harness rather than assuming it.