GnosisLLM Knowledge

Enterprise-grade knowledge loading, indexing, and semantic search library for Python.

Features

• Semantic Search: Vector-based similarity search using OpenAI embeddings
• Hybrid Search: Combine semantic and keyword (BM25) search for best results
• Agentic Search: AI-powered search with reasoning and natural language answers
• Agentic Memory: Conversational memory with automatic fact extraction
• Multiple Loaders: Load content from websites, sitemaps, and files
• Intelligent Chunking: Sentence-aware text splitting with configurable overlap
• OpenSearch Backend: Production-ready with k-NN vector search
• Multi-Tenancy: Index isolation for complete tenant separation (tenant-agnostic library)
• Event-Driven: Observer pattern for progress tracking and monitoring
• SOLID Architecture: Clean, maintainable, and extensible codebase

Installation

pip install gnosisllm-knowledge

# With OpenSearch backend
pip install gnosisllm-knowledge[opensearch]

# With all optional dependencies
pip install gnosisllm-knowledge[all]

Quick Start (CLI)

# Install
pip install gnosisllm-knowledge

# Set OpenAI API key for embeddings
export OPENAI_API_KEY=sk-...

# Setup OpenSearch with ML model
gnosisllm-knowledge setup --host localhost --port 9200
# ✓ Created connector, model, pipelines, index
# Model ID: abc123  →  Add to .env: OPENSEARCH_MODEL_ID=abc123

export OPENSEARCH_MODEL_ID=abc123

# Load content from a sitemap
gnosisllm-knowledge load https://docs.example.com/sitemap.xml
# ✓ Loaded 247 documents (1,248 chunks) in 45.3s

# Search
gnosisllm-knowledge search "how to configure authentication"
# Found 42 results (23.4ms)
# 1. Authentication Guide (92.3%)
#    To configure authentication, set AUTH_PROVIDER...

# Interactive search mode
gnosisllm-knowledge search --interactive

Quick Start (Python API)

from gnosisllm_knowledge import Knowledge

# Create instance with OpenSearch backend
knowledge = Knowledge.from_opensearch(
    host="localhost",
    port=9200,
)

# Setup backend (creates indices)
await knowledge.setup()

# Load and index a sitemap
await knowledge.load(
    "https://docs.example.com/sitemap.xml",
    collection_id="docs",
)

# Search
results = await knowledge.search("how to configure authentication")
for item in results.items:
    print(f"{item.title}: {item.score}")

CLI Commands

Setup

Configure OpenSearch with neural search capabilities:

gnosisllm-knowledge setup [OPTIONS]

Options:
  --host        OpenSearch host (default: localhost)
  --port        OpenSearch port (default: 9200)
  --use-ssl     Enable SSL connection
  --force       Clean up existing resources first
  --no-hybrid   Skip hybrid search pipeline

Load

Load and index content from URLs or sitemaps:

gnosisllm-knowledge load <URL> [OPTIONS]

Options:
  --type         Source type: website, sitemap (auto-detects)
  --index        Target index name (e.g., knowledge-tenant-123)
  --collection-id Collection grouping ID
  --batch-size   Documents per batch (default: 100)
  --max-urls     Max URLs from sitemap (default: 1000)
  --dry-run      Preview without indexing

Multi-tenancy is achieved through index isolation. Use --index with tenant-specific names (e.g., --index knowledge-tenant-123).

Search

Search indexed content with multiple modes:

gnosisllm-knowledge search <QUERY> [OPTIONS]

Options:
  --mode         Search mode: semantic, keyword, hybrid, agentic
  --index        Index to search (e.g., knowledge-tenant-123)
  --limit        Max results (default: 5)
  --collection-ids Filter by collections (comma-separated)
  --json         Output as JSON for scripting
  --interactive  Interactive search session

Multi-tenancy is achieved through index isolation. Use --index with tenant-specific names.

Architecture

gnosisllm-knowledge/
├── api/                 # High-level Knowledge facade
├── core/
│   ├── domain/          # Document, SearchQuery, SearchResult models
│   ├── interfaces/      # Protocol definitions (IContentLoader, etc.)
│   ├── events/          # Event system for progress tracking
│   └── exceptions.py    # Exception hierarchy
├── loaders/             # Content loaders (website, sitemap)
├── fetchers/            # Content fetchers (HTTP, Neoreader)
├── chunking/            # Text chunking strategies
├── backends/
│   ├── opensearch/      # OpenSearch implementation
│   └── memory/          # In-memory backend for testing
└── services/            # Indexing and search orchestration

Search Modes

from gnosisllm_knowledge import SearchMode

# Semantic search (vector similarity)
results = await knowledge.search(query, mode=SearchMode.SEMANTIC)

# Keyword search (BM25)
results = await knowledge.search(query, mode=SearchMode.KEYWORD)

# Hybrid search (default - combines both)
results = await knowledge.search(query, mode=SearchMode.HYBRID)

Agentic Search

AI-powered search with reasoning and natural language answers using OpenSearch ML agents.

Requirements: OpenSearch 3.4+ for conversational memory support.

Setup

# 1. First run standard setup (creates embedding model)
gnosisllm-knowledge setup --port 9201

# 2. Setup agentic agents (creates LLM connector, VectorDBTool, MLModelTool, agents)
gnosisllm-knowledge agentic setup
# ✓ Flow Agent ID: abc123
# ✓ Conversational Agent ID: def456

# 3. Add agent IDs to environment
export OPENSEARCH_FLOW_AGENT_ID=abc123
export OPENSEARCH_CONVERSATIONAL_AGENT_ID=def456

Usage

# Single-turn agentic search (uses flow agent)
gnosisllm-knowledge search --mode agentic "What is Typer?"

# Interactive multi-turn chat (uses conversational agent with memory)
gnosisllm-knowledge agentic chat
# You: What is Typer?
# Assistant: Typer is a library for building CLI applications...
# You: What did you just say about it?
# Assistant: I told you that Typer is a library for building CLI...

How It Works

┌─────────────────────────────────────────────────────────────────────┐
│                         User Query                                   │
│                    "What is Typer?"                                  │
└─────────────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────────────┐
│                    OpenSearch ML Agent                               │
│              (Flow or Conversational)                                │
└─────────────────────────────────────────────────────────────────────┘
                              │
            ┌─────────────────┴─────────────────┐
            ▼                                   ▼
┌───────────────────────┐           ┌───────────────────────┐
│     VectorDBTool      │           │   Conversation Memory │
│  (Knowledge Search)   │           │   (Conversational     │
│                       │           │    Agent Only)        │
│  - Searches index     │           │                       │
│  - Returns context    │           │  - Stores Q&A pairs   │
│                       │           │  - Injects chat_history│
└───────────────────────┘           └───────────────────────┘
            │                                   │
            └─────────────────┬─────────────────┘
                              ▼
┌─────────────────────────────────────────────────────────────────────┐
│                      MLModelTool (answer_generator)                  │
│                                                                      │
│  Prompt Template:                                                    │
│  ┌────────────────────────────────────────────────────────────────┐ │
│  │ Context from knowledge base:                                    │ │
│  │ ${parameters.knowledge_search.output}                           │ │
│  │                                                                 │ │
│  │ Previous conversation:        ← Only for conversational agent  │ │
│  │ ${parameters.chat_history:-}                                    │ │
│  │                                                                 │ │
│  │ Question: ${parameters.question}                                │ │
│  └────────────────────────────────────────────────────────────────┘ │
└─────────────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────────────┐
│                      AI-Generated Answer                             │
│  "Typer is a library for building CLI applications in Python..."    │
└─────────────────────────────────────────────────────────────────────┘

Agent Types

Agent	Type	Use Case	Memory
Flow	flow	Fast single-turn RAG, API calls	No
Conversational	conversational_flow	Multi-turn dialogue, chat	Yes

Key Configuration

The conversational agent requires these settings for memory to work:

# In agent registration (setup.py)
agent_body = {
    "type": "conversational_flow",
    "app_type": "rag",  # Required for memory injection
    "llm": {
        "model_id": llm_model_id,
        "parameters": {
            "message_history_limit": 10,  # Include last N messages
        },
    },
    "memory": {"type": "conversation_index"},
}

# MLModelTool prompt must include:
# ${parameters.chat_history:-}  ← Receives conversation history

Multi-Tenancy

This library is tenant-agnostic. Multi-tenancy is achieved through index isolation - each tenant gets their own OpenSearch index.

# The calling application (e.g., API) constructs tenant-specific index names
index_name = f"knowledge-{account_id}"

# Create Knowledge instance for the tenant
knowledge = Knowledge.from_opensearch(
    host="localhost",
    port=9200,
    index_prefix=index_name,  # knowledge-tenant-123
)

# Load content to tenant's isolated index
await knowledge.load(
    source="https://docs.example.com/sitemap.xml",
    collection_id="docs",
)

# Search within tenant's index (no account_id filter needed)
results = await knowledge.search(
    "query",
    collection_ids=["docs"],
)

Note: For audit purposes, you can store account_id in document metadata:

await knowledge.load(
    source="https://docs.example.com/sitemap.xml",
    document_defaults={"metadata": {"account_id": "tenant-123"}},
)

Agentic Memory

Conversational memory with automatic fact extraction using OpenSearch's ML Memory plugin.

# Setup memory connectors
gnosisllm-knowledge memory setup --openai-key sk-...

# Create container and store conversations
gnosisllm-knowledge memory container create my-memory
gnosisllm-knowledge memory store <container-id> --file messages.json --user-id alice
gnosisllm-knowledge memory recall <container-id> "user preferences" --user-id alice

from gnosisllm_knowledge import Memory, MemoryStrategy, StrategyConfig, Message

memory = Memory.from_env()

# Create container with strategies
container = await memory.create_container(
    name="agent-memory",
    strategies=[
        StrategyConfig(type=MemoryStrategy.SEMANTIC, namespace=["user_id"]),
    ],
)

# Store conversation with fact extraction
await memory.store(
    container_id=container.id,
    messages=[Message(role="user", content="I prefer dark mode")],
    user_id="alice",
    infer=True,
)

# Recall memories
result = await memory.recall(container.id, "preferences", user_id="alice")

See docs/memory.md for full documentation.

Event Tracking

from gnosisllm_knowledge import EventType

# Subscribe to events
@knowledge.events.on(EventType.DOCUMENT_INDEXED)
def on_indexed(event):
    print(f"Indexed: {event.document_id}")

@knowledge.events.on(EventType.BATCH_COMPLETED)
def on_batch(event):
    print(f"Batch complete: {event.documents_indexed} docs")

Configuration

from gnosisllm_knowledge import OpenSearchConfig

# From environment variables
config = OpenSearchConfig.from_env()

# Explicit configuration
config = OpenSearchConfig(
    host="search.example.com",
    port=443,
    use_ssl=True,
    username="admin",
    password="secret",
    embedding_model="text-embedding-3-small",
    embedding_dimension=1536,
)

knowledge = Knowledge.from_opensearch(config=config)

Requirements

• Python 3.11+
• OpenSearch 2.0+ (for production use)
• OpenSearch 3.4+ (for agentic search with conversation memory)

License

MIT