ChaosCypher Core

Standalone AI knowledge graph library using Hexagonal Architecture (Ports & Adapters)

Quick Start

Zero-boilerplate facade (no database, no setup)

from chaoscypher_core import ChaosCypher

# Configure once (optional — defaults to Ollama, env vars auto-detected)
ChaosCypher.configure(provider="openai", api_key="sk-...")

# Extract entities from any document
result = ChaosCypher.extract_sync("paper.pdf")
print(f"{len(result.entities)} entities, {len(result.relationships)} relationships")

# Chat with an LLM
response = ChaosCypher.chat_sync("Explain knowledge graphs")
print(response.content)

# Embed text (single or batch)
single = ChaosCypher.embed_sync("quantum entanglement")
batch = ChaosCypher.embed_batch_sync(["text one", "text two"])

# Process documents into a knowledge graph
result = ChaosCypher.add_document_sync("paper.pdf", database="demo")
results = ChaosCypher.add_documents_sync(["doc1.pdf", "doc2.pdf"])

# Extract from raw text (no file needed)
result = await ChaosCypher.extract(text="Alice knows Bob. Bob works at Acme.")
chunks = await ChaosCypher.chunk(text="Long document content...")

A short alias CC is also available: from chaoscypher_core import CC.

Every method has an async variant (extract, chat, embed, embed_batch, search, add_document, add_documents, chunk) and a sync variant with a _sync suffix.

Build a knowledge graph (with persistent storage)

from chaoscypher_core import Engine

# Inline configuration — same kwargs as ChaosCypher.configure()
with Engine(database="demo", provider="openai", api_key="sk-...") as engine:
    # Quick graph building with auto-created templates
    alice = engine.add_node("Person", "Alice", properties={"role": "Engineer"})
    bob = engine.add_node("Person", "Bob")
    engine.add_edge("knows", alice, bob)

    # Process a document into the graph
    result = engine.add_document_sync("paper.pdf")
    print(f"Extracted {len(result.nodes)} nodes")

    # Search the graph (sync and async)
    results = engine.search_sync("quantum entanglement")
    results = await engine.search("quantum entanglement", mode="semantic")

    # Chat and embed (sync wrappers for scripts/notebooks)
    response = engine.chat_sync("Summarize this graph")
    embedding = engine.embed_sync("quantum physics")
    embeddings = engine.batch_embed_sync(["text one", "text two"])

    stats = engine.get_stats()
    print(f"Graph: {stats.nodes} nodes, {stats.edges} edges")

Engine also accepts data_dir= for an explicit path instead of a database name.

Overview

A reusable, framework-agnostic library providing core AI-powered knowledge graph capabilities:

• Entity Extraction: AI-powered extraction from documents with template matching and deduplication
• Document Loading: Auto-detect file type (PDF, text, HTML, RST, DOCX, XLSX, PPTX, EPUB, CSV, JSON, JSONL, audio, video, image, archives)
• RAG Pipeline: Chunking, vector embeddings, and hybrid search (keyword + semantic)
• Graph Management: Node, edge, and template CRUD with search index synchronization
• Graph Analytics: Centrality, clustering, path analysis, community detection
• Workflow Execution: LangGraph-based workflow engine with tool calling
• LLM Integration: Ollama, OpenAI, Anthropic, and Gemini providers

Architecture: Hexagonal (Ports & Adapters)

┌─────────────────────────────────────────────────────────────┐
│                     PORTS (Interfaces)                       │
│  ports/  - Python Protocols defining contracts               │
│  - GraphRepositoryProtocol, SearchRepositoryProtocol, etc.   │
└─────────────────────────────────────────────────────────────┘
                              ▲
                              │
┌─────────────────────────────┼───────────────────────────────┐
│                        CORE (Business Logic)                 │
│  services/  - Domain-organized business rules                │
│  - graph/, search/, sources/, workflows/, chat/, export/     │
│                                                              │
│  models.py  - Pure Pydantic DTOs (no ORM dependencies)       │
│  exceptions.py - Domain-specific exceptions                  │
└──────────────────────────────────────────────────────────────┘
                              ▲
                              │
┌─────────────────────────────┼───────────────────────────────┐
│                     ADAPTERS (Implementations)               │
│  adapters/sqlite/        - SQLite storage (default)          │
│  adapters/llm/           - LLM providers (Ollama, OpenAI, …) │
│  adapters/sqlite/repos/  - Graph, search, and extraction     │
│                            repositories (sqlite-vec + FTS5)  │
└──────────────────────────────────────────────────────────────┘

Why Hexagonal?

1. Framework Independence: No FastAPI, SQLModel, or Docker dependencies in core
2. Testability: Easy to mock protocol interfaces for unit testing
3. Portability: Embeddable in web apps, CLIs, Jupyter notebooks, or scripts
4. Dual Deployment: Supports async (Docker + Valkey) and sync (standalone) modes

Import Strategy

Namespace class (recommended for most users):

from chaoscypher_core import ChaosCypher  # or CC

# One import gives you IDE autocomplete for every method
result = ChaosCypher.extract_sync("paper.pdf")
response = ChaosCypher.chat_sync("Explain knowledge graphs")
embedding = ChaosCypher.embed_sync("quantum entanglement")
batch = ChaosCypher.embed_batch_sync(["text one", "text two"])
results = ChaosCypher.search_sync("quantum", database="demo")

Granular imports (organized by tier):

from chaoscypher_core import (
    # --- Primary API (start here) ---
    ChaosCypher, CC, Engine,
    # Async convenience functions
    extract, chat, embed, embed_batch, search, chunk,
    add_document, add_documents,
    # Sync convenience functions
    extract_sync, chat_sync, embed_sync, embed_batch_sync, search_sync, chunk_sync,
    add_document_sync, add_documents_sync,

    # --- Configuration ---
    EngineSettings, LLMSettings,

    # --- Models (input DTOs) ---
    NodeCreate, NodeUpdate, NodePosition,
    EdgeCreate, EdgeUpdate,
    TemplateCreate, TemplateUpdate,
    PropertyDefinition, PropertyType,
    AnalysisDepth, SearchMode, ProgressCallback, ProgressStage,

    # --- Models (output DTOs) ---
    Node, Edge, Template,
    ExtractionResult, ProcessingResult, LLMChatResponse,
    EmbedResult, BatchEmbedResult, EngineSearchResult,
    ChunksResult, ChunkingResult, IndexingResult, RebuildResult,
    PaginatedResult, DatabaseStats, DatabaseInfo,
    HealthReport, HealthResult, TokenUsage, ToolResult, SourceStatus,

    # --- Services (for Engine direct usage) ---
    NodeService, EdgeService, TemplateService,
    IndexingService, SearchService,
    ExtractionService, SourceCommitService,
    ChunkingService, Loaders,

    # --- Advanced (adapters, protocols, factories) ---
    LLMProvider, BaseLLMProvider, ProviderFactory, SqliteAdapter,
    create_embedding_provider, ToolExecutionContext, generate_id,
    # Protocols (for custom adapter authors)
    GraphRepositoryProtocol, SearchRepositoryProtocol,
    ChunkingProtocol, IndexingProtocol, EmbeddingProviderProtocol,
    SourceStorageProtocol, CitationStorageProtocol,
)

The primary API, models, services, and the protocols shown above are all top-level imports. Some granular storage protocols (workflows, chats, tools, triggers, LLM metrics) are not re-exported and require a deep import from their chaoscypher_core.ports.storage_* module — see the Storage Protocols table below.

Directory Structure

chaoscypher_core/
├── __init__.py              # Public API exports
├── models.py                # Pure Pydantic DTOs
├── settings.py              # EngineSettings configuration
├── exceptions.py            # Exception hierarchy
├── bootstrap.py             # Engine class (single entry point)
│
├── ports/                   # PORTS: Protocol definitions (contracts)
│   ├── graph.py             # GraphRepositoryProtocol
│   ├── search.py            # SearchRepositoryProtocol
│   ├── chunk.py             # ChunkingProtocol
│   ├── index.py             # IndexingProtocol
│   ├── embedding.py         # EmbeddingProviderProtocol
│   ├── llm.py               # LLM provider protocols
│   ├── db.py                # DatabaseProtocol
│   └── storage_*.py         # Granular storage protocols, one file per concern
│                            # (storage_sources, storage_chats, storage_tools,
│                            #  storage_triggers, storage_workflows, …)
│
├── services/                # CORE: Business logic (domain-organized)
│   ├── graph/               # Graph CRUD (NodeService, EdgeService, TemplateService)
│   │   ├── engine/          # Analytics (centrality, clustering, paths)
│   │   └── management/      # Node, edge, template, source management
│   ├── search/              # RAG pipeline
│   │   └── engine/          # IndexingService, SearchService
│   ├── sources/             # Document processing
│   │   ├── engine/          # Extraction, commit, deduplication
│   │   │   └── extraction/utils/post_extraction.py  # Shared structural-filter + type-normalize (Cortex/CLI/MCP parity)
│   │   ├── loaders/         # File type loaders (PDF, text, HTML, RST, DOCX, XLSX, PPTX, EPUB, CSV, JSON, JSONL, audio, video, image)
│   │   ├── normalizer/      # Content cleaning (cleaners return CleanerResult; OCRCleaner overrides applies_to)
│   │   └── management/      # Source CRUD
│   ├── workflows/           # Workflow orchestration
│   │   ├── engine/          # LangGraph executor
│   │   ├── management/      # Workflow CRUD
│   │   ├── tools/           # Tool plugins
│   │   └── triggers/        # Event triggers
│   ├── chat/                # Conversational AI
│   ├── export/              # Graph data export
│   ├── quality/             # Entity quality scoring + per-source quality counters (counters.py)
│   ├── compose/             # Graph merge/resolve
│   ├── lexicon/             # Lexicon Hub client, auth, credential storage
│   ├── presets/             # Preset configurations
│   └── package/             # Import/export packages
│
├── adapters/                # ADAPTERS: External integrations
│   ├── sqlite/              # SQLite storage adapter (default)
│   │   ├── adapter.py       # SqliteAdapter (implements all storage protocols)
│   │   ├── models.py        # SQLModel entity definitions
│   │   ├── mixins/          # Protocol implementations (one focused mixin per concern)
│   │   └── repos/           # Repository implementations
│   │       ├── graph/       # GraphRepository (SQLite-backed)
│   │       ├── search.py    # SearchRepository (sqlite-vec + FTS5)
│   │       └── extraction.py  # ExtractionRepository
│   ├── llm/                 # LLM provider system
│   │   ├── factory.py       # ProviderFactory (cached provider creation)
│   │   ├── provider.py      # LLMProvider (queue-free direct access)
│   │   └── providers/       # Ollama, OpenAI, Anthropic, Gemini
│   └── web/                 # Web/HTTP integrations
│
├── plugins/                 # Plugin system (base classes, discovery, registry)
├── utils/                   # Utilities (ID generation, chunking, logging)
│   ├── encoding.py          # detect_encoding(path) — text-shaped loader pre-read
│   └── normalization_default.py  # resolve_normalization_default(filename) — CSV/JSON default off
└── data/                    # Static data files

API Reference

ChaosCypher Facade

Static namespace class for one-import convenience. No instantiation needed.

extract() and chunk() accept text= for explicit text input, skipping file detection.

embed_batch() / embed_batch_sync() always return BatchEmbedResult (no union type).

Engine

Persistent graph engine with full service access. Accepts the same provider=, api_key=, and configuration kwargs as ChaosCypher.configure().

Engine(database="mydb", provider="openai", api_key="sk-...")
Engine(data_dir="./data/databases/mydb")            # explicit path
Engine(database="mydb", settings=my_engine_settings) # pre-built settings

Convenience methods (return Pydantic models):

Sync wrappers (for scripts and notebooks):

search_sync, chat_sync, embed_sync, batch_embed_sync, add_document_sync, add_documents_sync, process_document_sync

Advanced: Direct Service Access

For fine-grained control, Engine exposes the underlying services and models. Use TemplateCreate, NodeCreate, and EdgeCreate DTOs when you need full control over IDs, properties, and template configuration:

from chaoscypher_core import Engine, NodeCreate, PropertyDefinition, TemplateCreate

with Engine(database="demo") as engine:
    # Create a template with property definitions
    t = engine.create_template(TemplateCreate(
        name="Person",
        template_type="node",
        properties=[
            PropertyDefinition(name="role", display_name="Role", property_type="string"),
        ],
    ))

    # Create nodes with explicit template ID
    alice = engine.create_node(NodeCreate(
        template_id=t.id, label="Alice", properties={"role": "Engineer"},
    ))
    bob = engine.create_node(NodeCreate(
        template_id=t.id, label="Bob", properties={"role": "Designer"},
    ))

    # Direct service access (returns dicts, not models)
    nodes = engine.node_service.list_nodes()
    templates = engine.template_service.list_templates()

Storage Protocols

All storage protocols are defined in chaoscypher_core.ports and return dict[str, Any]. Protocols without an import path below are top-level exports (from chaoscypher_core import ...); the rest require a deep import from the listed module:

The default implementation for all protocols is SqliteAdapter (via focused per-concern mixins — one module per storage concern, following the Interface Segregation Principle).

Plugin Protocols

In addition to storage protocols, the core library exposes plugin protocols for extension points with a shared metadata contract:

• PluginMetadata (plugins/base.py) — unified Pydantic descriptor for all plugin types (loaders, tools, cleaners, archive handlers, LLM providers). Carries plugin_id / name / description / version / priority / applies_to / origin / tags.
• CleanerProtocol (services/sources/normalizer/cleaners/base.py) — clean(content, metadata) -> CleanerResult carrying (content, ops, lines_removed, paragraphs_deduplicated, chars_removed). Tuple unpacking still works via CleanerResult.__iter__ for back-compat. Cleaners may optionally implement applies_to(metadata) -> bool to gate themselves by document metadata (OCRCleaner uses this to fire only on OCR-derived content).
• ArchiveHandler (services/sources/loaders/archive/handlers/base.py) — carries metadata, a can_handle() -> int specificity score, and find_root() for nested-docs discovery.

Testing

Protocol-based design enables clean mocking. The core protocols are top-level exports (the granular storage protocols import from chaoscypher_core.ports.storage_* — see Storage Protocols above):

from unittest.mock import Mock
from chaoscypher_core import ExtractionService, GraphRepositoryProtocol

def test_extraction():
    service = ExtractionService(
        graph_repository=Mock(spec=GraphRepositoryProtocol),
        llm_provider=Mock(),
        settings=Mock(),
        embedding_service=None,  # required keyword-only; None disables semantic dedup
    )
    # Test business logic with mocked dependencies

Design Principles

• KISS: Simple, focused services with clear responsibilities
• DRY: Shared logic in utils/, protocol-based reuse
• SOLID: Interface segregation, dependency inversion, single responsibility
• No Technical Debt: Clean breaks over gradual migrations, delete dead code immediately

Documentation

For complete documentation including tutorials, API reference, and architecture guides, see https://chaoscypher.com or packages/docs/docs/developer-guide/ in the repository.

License

See main project LICENSE