MaMAI 4.3.0

MaMAI v4 è una libreria RAG stateless basata su LangChain con context esplicito e storage dinamico.

MaMAI v4.0.0

MaMAI is a stateless Python RAG library.

It is designed for service-oriented systems where the caller owns state and storage decisions.

Core principles

• No internal db.json or hidden persistence.
• No implicit global runtime configuration.
• Every operation receives a full ExecutionContext.
• Vector store backend and storage location are selected at runtime by the caller.

Installation

python3 -m venv .venv
. .venv/bin/activate
pip install -U pip
pip install -e .

Quick start

from Mama import (
    ExecutionContext,
    LlmConfig,
    RetrievalConfig,
    StorageConfig,
    default_runtime,
    ingest_pdf_directory,
    query,
)

ctx = ExecutionContext(
    tenant_id="acme",
    user_id="user-1",
    session_id="session-1",
    kb_id="kb-support",
    llm=LlmConfig(provider="dummy", model="dummy"),
    retrieval=RetrievalConfig(search_type="similarity", k=3),
    storage=StorageConfig(
        backend="faiss_local",
        params={
            "base_path": "./data/vectorstores",
            "embeddings_backend": "simple",
        },
    ),
)

runtime = default_runtime()
ingest_pdf_directory(ctx, "./docs_to_index", runtime)
result = query(
    ctx,
    "What is this document about?",
    runtime,
    prompt=(
        "You are the KB assistant. Follow global KB rules.\n"
        "Apply instructions from the active knowledge branch."
    ),
)

print(result.answer)
print(result.documents)

# Caller-owned state persistence
ctx.conversation = result.updated_conversation_state

ExecutionContext

ExecutionContext is the contract between your wrapper/service and MaMAI core.

Required identity fields:

• tenant_id
• user_id
• session_id
• kb_id

Config sections:

• llm: provider, model, parameters
• retrieval: search type and k
• storage: backend + dynamic backend params
• conversation: in-memory message state passed in/out

Per-query prompt input (required by query(...)):

• prompt: fully resolved prompt string for the current query

You can build it programmatically or from JSON payload with:

from Mama.config import build_execution_context
ctx = build_execution_context(payload)

Identity Model (tenant_id, user_id, session_id, kb_id)

These 4 fields are not interchangeable. They model different scopes.

tenant_id (organization boundary)

• Purpose: hard isolation boundary between customers/workspaces.
• Typical value: company/workspace slug or UUID (acme, tenant-42).
• Effect in MaMAI:
  • used by storage naming for faiss_local (<base_path>/<tenant_id>/<kb_id>)
  • should also be propagated to metadata and external stores.
• Rule: never mix data from different tenants in one index.

user_id (end-user identity)

• Purpose: actor identity (who is asking / who ingested).
• Typical value: authenticated user id (user-123, email hash, UUID).
• Effect in MaMAI:
  • not used for index partitioning by default
  • available for metadata, auditing, policy decisions in wrappers.
• Rule: treat as identity/audit field, not as storage key unless your design needs per-user KBs.

session_id (conversation instance)

• Purpose: one chat/conversation thread.
• Typical value: request/session UUID (sess-2026-...).
• Effect in MaMAI:
  • identifies the conversation memory passed in/out via ctx.conversation
  • does not create separate vector indexes by default.
• Rule: rotate session_id per new chat thread; persist chat state in your own store.

kb_id (knowledge base identity)

• Purpose: logical knowledge corpus id.
• Typical value: domain/corpus key (kb-support, kb-legal-v2).
• Effect in MaMAI:
  • primary key for vector index selection together with tenant_id
  • same kb_id means same index namespace (inside tenant).
• Rule: change kb_id when you want a different corpus/index.

Practical composition

Think in scopes:

• Tenant scope: tenant_id
• Corpus scope: kb_id
• User scope: user_id
• Conversation scope: session_id

For faiss_local, effective index path is:

• <base_path>/<tenant_id>/<kb_id>

So:

• changing session_id does not change index
• changing user_id does not change index (unless you encode it in kb_id)
• changing kb_id does change index
• changing tenant_id does change isolation boundary and index location

Recommended patterns

Shared team KB + per-user chats:

• same tenant_id + kb_id
• different user_id/session_id

Versioned KB rollout:

• kb_id=kb-support-v1, then kb-support-v2
• run A/B or migration safely

Per-user private KB:

• set kb_id like kb-user-<user_id> (or explicit mapping in wrapper)

Anti-patterns to avoid

• Using session_id as kb_id accidentally (creates fragmented indexes).
• Reusing one global tenant_id for all customers (breaks isolation).
• Expecting user_id alone to separate vector stores (it does not by default).
• Keeping kb_id constant across unrelated corpora.

Caller-defined vector store storage

MaMAI does not decide where your index is stored. The caller does.

faiss_local

Persistent FAISS index on caller-provided filesystem location.

Supported params:

• base_path: root directory; final path is <base_path>/<tenant_id>/<kb_id>
• index_path: full explicit path (overrides base_path logic)
• embeddings_backend: simple, hf, or auto

Example with explicit per-request storage path:

ctx.storage = StorageConfig(
    backend="faiss_local",
    params={
        "index_path": "/mnt/customer-a/rag/kb-support-v1",
        "embeddings_backend": "simple",
    },
)

faiss_memory

In-memory FAISS store (non-persistent), useful for ephemeral jobs and tests.

ctx.storage = StorageConfig(
    backend="faiss_memory",
    params={"embeddings_backend": "simple"},
)

Register your own vector store backend

You can plug a custom backend (Qdrant, Pinecone, Weaviate, Milvus, custom S3 adapter).

from Mama.ports import RuntimePorts
from Mama.runtime import DefaultEmbeddingsProvider, DefaultLlmProvider
from Mama.vectorstores import DynamicVectorStoreFactory

factory = DynamicVectorStoreFactory()

# builder signature: (ctx, embeddings) -> VectorStorePort
factory.register_backend("qdrant", my_qdrant_builder)

runtime = RuntimePorts(
    llm_provider=DefaultLlmProvider(),
    embeddings_provider=DefaultEmbeddingsProvider(),
    vector_store_factory=factory,
)

ctx.storage.backend = "qdrant"
ctx.storage.params = {
    "url": "https://qdrant.example.com",
    "api_key": "...",
    "collection": "kb-support",
    "tenant": ctx.tenant_id,
}

State ownership model

MaMAI core is stateless by design:

• it reads conversation state from ctx.conversation
• it returns updated state in QueryResult.updated_conversation_state
• it does not persist sessions/history internally

Typical wrapper responsibilities:

• load context from request + tenant config
• call MaMAI
• persist updated conversation state in your DB/cache
• enforce auth, rate limits, and policy

Main modules

• Mama/models.py: data contracts (ExecutionContext, QueryResult, IngestResult)
• Mama/ports.py: extension contracts for runtime providers
• Mama/runtime.py: default runtime providers
• Mama/vectorstores.py: dynamic vector store factory and FAISS adapters
• Mama/ingestion.py: stateless ingestion use cases
• Mama/query.py: stateless RAG query pipeline
• Mama/admin.py: document listing and vector store stats
• Mama/webscraper.py: URL crawl/load/ingest helpers

Testing

Run smoke test:

make test

Run web demo:

make web

Health endpoint:

curl http://127.0.0.1:5000/health

Notes

The sample tests generate a minimal PDF file and may show pypdf parsing warnings. If the test ends with TEST OK, the pipeline worked as expected.