LLM Parameter Filtering#

ScrapeGraphAI's _create_llm() method in AbstractGraph constructs LLM client instances from a user-supplied llm config dict. Several keys in that dict are ScrapeGraphAI-internal — consumed by the framework and meaningless (or actively harmful) to underlying LangChain/provider constructors. These keys must be removed before the dict is splatted into a provider class or init_chat_model().

Internal Parameters That Are Filtered#

The model_tokens Bug and Fix (PR #1100)#

The model_tokens key was a known leakage point. On the plain llm-config path, _create_llm() reads the value into self.model_token 1 but, before the fix, never removed it from llm_params. When init_chat_model(**llm_params) was then called, model_tokens was forwarded to the underlying model client, producing:

TypeError: Completions.create() got an unexpected keyword argument 'model_tokens'. Did you mean 'max_tokens'?

PR #1100 fixes this with a single pop and explanatory comment immediately after self.model_token is set:

# Consumed by ScrapeGraphAI; must not be forwarded to the model client.
llm_params.pop("model_tokens", None)

This mirrors the behavior on the model_instance path 2, which returns early before any keys can reach the provider.

How Filtering Is Structured (No Central Utility)#

There is no centralized filter list or utility function. Filtering is ad-hoc inside _create_llm(), with each internal key popped at the appropriate point in the method's flow:

1. rate_limit — popped early at the top of the method 3
2. model_tokens — popped after token count is resolved (after line 222)
3. model_provider — popped in the custom-provider branch before constructing the wrapper 4

The model_instance key is effectively filtered by the early-return at 2 — if model_instance is present, the method returns immediately and llm_params is never passed to any constructor.

Implications for Adding New Providers#

Any new custom provider branch added to _create_llm() must:

1. Pop model_provider (already done at line 243 before the branch).
2. Pop model_tokens before calling ProviderClass(**llm_params).
3. Pop any other ScrapeGraphAI-specific keys present in llm_params at that point.

Failure to do so will cause a TypeError at runtime when the provider constructor receives an unexpected keyword argument — exactly the bug described in PR #1100 5.

Key References#

• abstract_graph.py — _create_llm() method
• PR #1100 — fix: pop model_tokens so it is not forwarded to the model client