Vector Search

Vector search retrieves context by semantic similarity rather than exact keywords — useful when the wording in outer_context differs from your query.

Install

Vector search needs optional dependencies (FAISS + embeddings):

pip install 'mmar-carl[vector-search]'

This pulls in faiss-cpu, fastembed, and numpy. Substring search always works without these.

Configure

from mmar_carl import ContextSearchConfig, ReasoningChain

search_config = ContextSearchConfig(
    strategy="vector",
    embedding_model="all-MiniLM-L6-v2",   # optional; a sensible default is used otherwise
    vector_config={
        "index_type": "flat",        # "flat" (default) or "ivf" for large datasets
        "similarity_threshold": 0.7, # minimum similarity score, 0–1 (default 0.7)
        "max_results": 5,            # max snippets per query (default 5)
    },
)

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

vector_config key	Default	Purpose
index_type	"flat"	FAISS index type; "ivf" scales to large corpora.
similarity_threshold	0.7	Drop matches below this score.
max_results	5	Max snippets returned per query.

Substring vs vector

	Substring	Vector
Dependencies	none	faiss-cpu, fastembed, numpy
Matching	exact keywords	semantic similarity
Speed	fastest	slower (embeds text)
Best for	codes, IDs, exact terms	paraphrased / conceptual matches

Mix both in one step with per-query overrides.