[FEATUREE] Allow passing extraction model via ExtractionConfig instead of env var

[voidwisp]

The only ways to set the extraction LLM model are:

1. Set EXTRACTION_MODEL env var before the config singleton lazily initializes
2. Directly set GraphRAGConfig.extraction_llm (undocumented, reaches into internals)

For multi-tenant pipelines where each tenant uses a different model, neither approach is clean. Setting env vars as a side effect before calling extract_and_build() is fragile — it relies on the singleton not having been initialized yet, and mutating global state between tenants requires knowing that
_extraction_llm gets reset (it doesn't, so switching models between tenants is actually broken).

Suggestion:

Accept extraction_llm (model ID string or LLM instance) as a parameter on ExtractionConfig or IndexingConfig, and have it take precedence over the env var / default:

extraction_config = ExtractionConfig(
extraction_llm='us.anthropic.claude-sonnet-4-5-20250929-v1:0',
preferred_entity_classifications=classifications,
)

This keeps configuration explicit and co-located, avoids global state mutation, and supports multi-tenant use cases naturally.

Current workaround:

os.environ['EXTRACTION_MODEL'] = tenant.model

[iansrobinson]

Hi @voidwisp

Thank you - good suggestion and good insight into how some applications might use the multi-tenancy feature. We can add this for extraction.

I imagine something similar might apply on the querying side. Currently you can set either GraphRAGConfig.response_llm or the RESPONSE_MODEL environment variable. We might also allow you to override on a per-tenant basis like this:

query_engine = LexicalGraphQueryEngine.for_traversal_based_search(
        graph_store, 
        vector_store,
        tenant_id='aoss',
        response_llm='us.anthropic.claude-sonnet-4-5-20250929-v1:0'
 )