AMSDAL ML

Machine learning plugin for the AMSDAL Framework, providing embeddings, vector search, semantic retrieval, and AI agents with support for OpenAI models.

Features

• Vector Embeddings: Generate and store embeddings for any AMSDAL model with automatic chunking
• Semantic Search: Query your data using natural language with tag-based filtering
• AI Agents: Build Q&A systems with streaming support and citation tracking
• Async-First: Optimized for high-performance async operations
• MCP Integration: Expose and consume tools via Model Context Protocol (stdio/HTTP)
• File Attachments: Process and embed documents with built-in loaders
• Extensible: Abstract base classes for custom models, retrievers, and ingesters

Installation

pip install amsdal-ml

Requirements

• Python 3.11 or higher
• AMSDAL Framework 0.5.6+
• OpenAI API key (for default implementations)

Quick Start

1. Configuration

Create a .env file in your project root:

AMSDAL_ML_OPENAI_API_KEY=sk-your-api-key-here
AMSDAL_ML_ASYNC_MODE=true
AMSDAL_ML_ML_MODEL_CLASS=amsdal_ml.ml_models.openai_model.OpenAIModel
AMSDAL_ML_ML_RETRIEVER_CLASS=amsdal_ml.ml_retrievers.openai_retriever.OpenAIRetriever
AMSDAL_ML_ML_INGESTING_CLASS=amsdal_ml.ml_ingesting.openai_ingesting.OpenAIIngesting

Create a config.yml for AMSDAL connections:

application_name: my-ml-app
async_mode: true
connections:
  - name: sqlite_state
    backend: sqlite-state-async
    credentials:
      - db_path: ./warehouse/state.sqlite3
      - check_same_thread: false
  - name: lock
    backend: amsdal_data.lock.implementations.thread_lock.ThreadLock
resources_config:
  repository:
    default: sqlite_state
  lock: lock

2. Generate Embeddings

from amsdal_ml.ml_ingesting.openai_ingesting import OpenAIIngesting
from amsdal_ml.ml_config import ml_config

# Initialize ingesting
ingester = OpenAIIngesting(
    model=MyModel,
    embedding_field='embedding',
)

# Generate embeddings for an instance
instance = MyModel(content='Your text here')
embeddings = await ingester.agenerate_embeddings(instance)
await ingester.asave(embeddings, instance)

3. Semantic Search

from amsdal_ml.ml_retrievers.openai_retriever import OpenAIRetriever

retriever = OpenAIRetriever()

# Search for relevant content
results = await retriever.asimilarity_search(
    query='What is machine learning?',
    k=5,
    include_tags=['documentation']
)

for chunk in results:
    print(f'{chunk.object_class}:{chunk.object_id} - {chunk.raw_text}')

4. Build an AI Agent

from amsdal_ml.agents.default_qa_agent import DefaultQAAgent

agent = DefaultQAAgent()

# Ask questions
output = await agent.arun('Explain vector embeddings')
print(output.answer)
print(f'Used tools: {output.used_tools}')

# Stream responses
async for chunk in agent.astream('What is semantic search?'):
    print(chunk, end='', flush=True)

5. Functional Calling Agent with Python Tools

from amsdal_ml.agents.functional_calling_agent import FunctionalCallingAgent
from amsdal_ml.agents.python_tool import PythonTool
from amsdal_ml.ml_models.openai_model import OpenAIModel

llm = OpenAIModel()
agent = FunctionalCallingAgent(model=llm, tools=[search_tool, render_tool])
result = await agent.arun(user_query="Find products with price > 100", history=[])

6. Natural Language Query Retriever

from amsdal_ml.ml_retrievers.query_retriever import NLQueryRetriever

retriever = NLQueryRetriever(llm=llm, queryset=Product.objects.all())
documents = await retriever.invoke("Show me red products", limit=10)

7. Document Ingestion Pipeline

from amsdal_ml.ml_ingesting import ModelIngester
from amsdal_ml.ml_ingesting.pipeline import DefaultIngestionPipeline
from amsdal_ml.ml_ingesting.loaders.pdf_loader import PdfLoader
from amsdal_ml.ml_ingesting.processors.text_cleaner import TextCleaner
from amsdal_ml.ml_ingesting.splitters.token_splitter import TokenSplitter
from amsdal_ml.ml_ingesting.embedders.openai_embedder import OpenAIEmbedder
from amsdal_ml.ml_ingesting.stores.embedding_data import EmbeddingDataStore

pipeline = DefaultIngestionPipeline(
    loader=PdfLoader(),  # Uses pypdf for PDF processing
    cleaner=TextCleaner(),
    splitter=TokenSplitter(max_tokens=800, overlap_tokens=80),
    embedder=OpenAIEmbedder(),
    store=EmbeddingDataStore(),
)

ingester = ModelIngester(
    pipeline=pipeline,
    base_tags=["document"],
    base_metadata={"source": "pdf"},
)

Architecture

Core Components

• MLModel: Abstract interface for LLM inference (invoke, stream, with attachments)
• MLIngesting: Generate text and embeddings from data objects with chunking
• MLRetriever: Semantic similarity search with tag-based filtering
• Agent: Q&A and task-oriented agents with streaming and citations
• EmbeddingModel: Database model storing 1536-dimensional vectors linked to source objects
• PythonTool: Tool for executing Python functions within agents
• FunctionalCallingAgent: Agent specialized in functional calling with configurable tools
• NLQueryRetriever: Retriever for natural language queries on AMSDAL querysets
• DefaultIngestionPipeline: Pipeline for document ingestion including loader, cleaner, splitter, embedder, and store
• ModelIngester: High-level ingester for processing models with customizable pipelines and metadata
• PdfLoader: Document loader using pypdf for PDF processing
• TextCleaner: Processor for cleaning and normalizing text
• TokenSplitter: Splitter for dividing text into chunks based on token count
• OpenAIEmbedder: Embedder for generating embeddings via OpenAI API
• EmbeddingDataStore: Store for saving embedding data linked to source objects
• MCP Server/Client: Expose retrievers as tools or consume external MCP services

Configuration

All settings are managed via MLConfig in .env:

# Model Configuration
AMSDAL_ML_LLM_MODEL_NAME=gpt-4o
AMSDAL_ML_LLM_TEMPERATURE=0.0
AMSDAL_ML_EMBED_MODEL_NAME=text-embedding-3-small

# Chunking Parameters
AMSDAL_ML_EMBED_MAX_DEPTH=2
AMSDAL_ML_EMBED_MAX_CHUNKS=10
AMSDAL_ML_EMBED_MAX_TOKENS_PER_CHUNK=800

# Retrieval Settings
AMSDAL_ML_RETRIEVER_DEFAULT_K=8

Development

Setup

# Install dependencies
pip install --upgrade uv hatch==1.14.2
hatch env create
hatch run sync

Testing

# Run all tests with coverage
hatch run cov

# Run specific tests
hatch run test tests/test_openai_model.py

# Watch mode
pytest tests/ -v

Code Quality

# Run all checks (style + typing)
hatch run all

# Format code
hatch run fmt

# Type checking
hatch run typing

AMSDAL CLI

# Generate a new model
amsdal generate model MyModel --format py

# Generate property
amsdal generate property --model MyModel embedding_field

# Generate transaction
amsdal generate transaction ProcessEmbeddings

# Generate hook
amsdal generate hook --model MyModel on_create

MCP Server

Run the retriever as an MCP server for integration with Claude Desktop or other MCP clients:

python -m amsdal_ml.mcp_server.server_retriever_stdio \
  --amsdal-config "$(echo '{"async_mode": true, ...}' | base64)"

The server exposes a search tool for semantic search in your knowledge base.

Development

Release Workflow

1. Develop on a feature branch, create PR to main — CI runs lint + tests
2. When ready to release, create a release/X.Y.Z branch, bump version in amsdal_ml/__about__.py, update CHANGELOG.md
3. Merge release/* to main — CD workflow automatically creates tag, builds, publishes to PyPI, and creates GitHub Release with changelog

See RELEASE.md for the full step-by-step guide.

License

See LICENSE.txt for the AMSDAL End User License Agreement.

For third-party dependency licenses, see amsdal_ml/Third-Party Materials - AMSDAL Dependencies - License Notices.md.

Links

• AMSDAL Framework
• Documentation
• Issue Tracker