crewai-infino

CrewAI integration for Infino — give your agents semantic (vector), full-text (BM25), hybrid, and SQL retrieval as tools over one store on object storage. No second vector DB, no separate metadata store, no client-side fusion: one Infino table answers all four ways, and the agent picks the right one at runtime.

pip install crewai-infino

Agent in minutes

Build and populate a table with InfinoIndex, then hand its tools to an agent:

import infino
from crewai import Agent, Task, Crew
from crewai_infino import InfinoIndex

# Your embedding model — any callables that turn text into vectors.
# (e.g. sentence-transformers; dim must match.)
from my_embeddings import embed_documents, embed_query  # -> list[list[float]] / list[float]

conn = infino.connect("./infino-data")
index = InfinoIndex.create(
    conn, "kb",
    embed_documents=embed_documents, embed_query=embed_query, dim=384,
)
index.add_texts(
    ["Reset the X200 by holding power for 10s.", "Error E-507 means a stale cache."],
    metadatas=[{"product": "X200"}, {"product": "X200"}],
)

analyst = Agent(
    role="Support Analyst",
    goal="Answer customer questions from the knowledge base",
    backstory="You retrieve before you answer.",
    tools=index.as_tools(),          # semantic + keyword + hybrid + SQL, over one store
)

crew = Crew(agents=[analyst], tasks=[
    Task(description="How do I fix error E-507 on the X200?",
         expected_output="A concise fix.", agent=analyst),
])
print(crew.kickoff())

The tools

index.as_tools() (or infino_tools(searcher)) returns four crewai.tools.BaseTools:

Attach individual tools to different agents, or the whole set to one. The package bundles no embedding model — you pass embed callables, so you keep full control of the model (and the examples stay key-free with a local one).

Knowledge

Use Infino as a CrewAI knowledge backend so a crew auto-grounds its answers — no explicit tool call. InfinoKnowledgeStorage stores knowledge chunks in one Infino table and retrieves them by vector similarity:

from crewai import Crew
from crewai.knowledge.source.string_knowledge_source import StringKnowledgeSource
from crewai_infino import InfinoKnowledgeStorage

storage = InfinoKnowledgeStorage(
    connection=conn, embed_documents=embed_documents,
    embed_query=embed_query, dim=384,
)
source = StringKnowledgeSource(content="Error E-507 means a stale cache.")
crew = Crew(agents=[...], tasks=[...],
            knowledge_sources=[source], knowledge_storage=storage)

Lower-level API

• InfinoIndex — create / open a table, add_texts, delete, as_tools().
• InfinoSearcher — the read path: .semantic(), .keyword(), .hybrid(), .sql() returning plain row dicts. The tools are thin wrappers over it.

The table schema matches langchain-infino, so a table built by one adapter is queryable by the other.

Object storage (S3 / Azure)

Everything runs over any infino.Connection, so local disk and cloud object storage differ only in the URI and storage_options you pass to infino.connect — the index, tools, and knowledge backend are unchanged. Keys are the standard object_store config strings (aws_* / azure_*); ambient credentials (IAM role, env vars) need no storage_options at all.

# Amazon S3 (or S3-compatible: set aws_endpoint, aws_allow_http for MinIO/R2).
conn = infino.connect("s3://bucket/prefix", storage_options={
    "aws_access_key_id": "...",
    "aws_secret_access_key": "...",
    "aws_region": "us-east-1",
})

# Azure Blob Storage.
conn = infino.connect("az://container/prefix", storage_options={
    "azure_storage_account_name": "...",
    "azure_storage_account_key": "...",
})

index = InfinoIndex.create(
    conn, "kb", embed_documents=embed_documents, embed_query=embed_query, dim=384,
)

Pass validate=True to connect to probe the store at connect time so bad credentials fail there rather than on first read.

Status

Tools and the Knowledge storage backend (InfinoKnowledgeStorage) are the shipping surface. CrewAI Memory (durable agent memory) is planned next, targeting CrewAI's RAGStorage seam. The knowledge backend implements CrewAI's BaseKnowledgeStorage, which is 1.x-only — hence the crewai>=1.0 floor.

Development

make install     # editable install with test + lint extras
make unit        # engine-free unit tests
make integration # full suite against the engine
make lint type   # ruff + mypy

Apache-2.0.