Context Extraction

Each LLM step can declare step_context_queries. For every query, CARL searches your outer_context and injects the matching snippets into that step’s prompt — so each step sees only the context it needs. This is the RAG-like extraction at the heart of CARL.

LLMStepDescription(
    number=1,
    title="Financial analysis",
    aim="Analyze financial performance.",
    step_context_queries=["revenue growth", "profit margins", "cost efficiency"],
)

Choosing a strategy

Configure search at the chain level with ContextSearchConfig:

Field	Type	Default	Purpose
strategy	"substring" | "vector"	"substring"	The default search strategy.
substring_config	dict | None	None	Options for substring search.
vector_config	dict | None	None	Options for vector search.
embedding_model	str | None	None	Embedding model for vector search.

• Substring (default) — fast, exact keyword matching, no extra dependencies.
• Vector — semantic similarity via FAISS embeddings; see vector search.

Substring options

from mmar_carl import ContextSearchConfig, ReasoningChain

search_config = ContextSearchConfig(
    strategy="substring",
    substring_config={
        "case_sensitive": False,      # default
        "min_word_length": 2,         # default
        "max_matches_per_query": 3,   # default
    },
)

chain = ReasoningChain(steps=steps, search_config=search_config)

Or with the builder: ChainBuilder().with_search_config(search_config).

Per-query overrides

Mix plain string queries and ContextQuery objects in the same step to override the strategy for individual queries:

from mmar_carl import ContextQuery

step_context_queries=[
    "EBITDA",                                    # uses the chain default
    ContextQuery(
        query="revenue trends",
        search_strategy="vector",
        search_config={"similarity_threshold": 0.8, "max_results": 3},
    ),
    ContextQuery(
        query="NET_INCOME",
        search_strategy="substring",
        search_config={"case_sensitive": True},
    ),
]

ContextQuery field	Type	Purpose
query	str	The query text.
search_strategy	"substring" | "vector" | None	Override for this query.
search_config	dict | None	Extra search options for this query.

See also

• Vector search — semantic similarity with FAISS.
• LLM steps — where step_context_queries live.