LLMemory Basic Usage

Installation

uv add llmemory
# or
pip install llmemory

Prerequisites:

• Python 3.10 or higher
• PostgreSQL 14+ (tested up to PostgreSQL 16)
• pgvector extension 0.5.0+
• OpenAI API key (or configure local embeddings)

Installing pgvector:

# Ubuntu/Debian
sudo apt-get install postgresql-16-pgvector

# macOS with Homebrew
brew install pgvector

# Or using CREATE EXTENSION in PostgreSQL:
psql -d your_database -c "CREATE EXTENSION IF NOT EXISTS vector;"

Verifying pgvector installation:

SELECT * FROM pg_extension WHERE extname = 'vector';
-- Should return one row if installed correctly

API Overview

This skill documents core llmemory operations:

• LLMemory - Main interface class
• DocumentType - Enum for document types
• SearchType - Enum for search modes
• ChunkingStrategy - Enum for chunking strategies
• add_document() - Add and process documents
• search() - Search for documents
• searchwithrouting() - Search with automatic query routing (detects answerable queries)
• searchwithdocuments() - Search and return results with document metadata
• list_documents() - List documents with pagination
• get_document() - Retrieve a document (owner-scoped)
• getdocumentchunks() - Get chunks with pagination (owner-scoped)
• getchunkcount() - Get number of chunks for a document (owner-scoped)
• deletedocument() / deletedocuments() - Delete documents (owner-scoped)
• get_statistics() - Get owner statistics
• db_manager - Access underlying database manager
• initialize() / close() - Lifecycle management

Quick Start

import asyncio
from llmemory import LLMemory, DocumentType, SearchType

async def main():
    # Initialize
    memory = LLMemory(
        connection_string="postgresql://localhost/mydb",
        openai_api_key="sk-..."
    )
    await memory.initialize()

    # Add a document
    result = await memory.add_document(
        owner_id="workspace-1",
        id_at_origin="user-123",
        document_name="example.txt",
        document_type=DocumentType.TEXT,
        content="Your document content here...",
        metadata={"category": "example"}
    )
    print(f"Created document with {result.chunks_created} chunks")

    # Search
    results = await memory.search(
        owner_id="workspace-1",
        query_text="your search query",
        search_type=SearchType.HYBRID,
        limit=5
    )
    for result in results:
        print(f"[{result.score:.3f}] {result.content[:80]}...")

    # Clean up
    await memory.close()

asyncio.run(main())

Complete API Documentation

LLMemory

Main interface for document operations.

Constructor:

LLMemory(
    connection_string: Optional[str] = None,
    openai_api_key: Optional[str] = None,
    config: Optional[LLMemoryConfig] = None,
    db_manager: Optional[AsyncDatabaseManager] = None
)

Parameters:

• connectionstring (str, optional): PostgreSQL connection URL (format: postgresql://user:pass@host:port/database). Ignored if dbmanager provided.
• openaiapikey (str, optional): OpenAI API key for embeddings. Can also be set via OPENAIAPIKEY environment variable.
• config (LLMemoryConfig, optional): Configuration object. Defaults to config from environment if not provided.
• db_manager (AsyncDatabaseManager, optional): Existing database manager from shared pool (for production apps with multiple services).

Raises:

• ConfigurationError: If neither connectionstring nor dbmanager provided, or if configuration is invalid.

Example:

from llmemory import LLMemory

# Simple initialization
memory = LLMemory(
    connection_string="postgresql://localhost/mydb",
    openai_api_key="sk-..."
)
await memory.initialize()

LLMemory.fromdbmanager()

Create instance from existing AsyncDatabaseManager (shared pool pattern).

Signature:

@classmethod
def from_db_manager(
    cls,
    db_manager: AsyncDatabaseManager,
    openai_api_key: Optional[str] = None,
    config: Optional[LLMemoryConfig] = None
) -> LLMemory

Parameters:

• db_manager (AsyncDatabaseManager, required): Existing database manager with schema already set
• openaiapikey (str, optional): OpenAI API key
• config (LLMemoryConfig, optional): Configuration object

Returns:

• LLMemory: Configured instance

Example:

from pgdbm import AsyncDatabaseManager, DatabaseConfig
from llmemory import LLMemory

# Create shared pool
config = DatabaseConfig(connection_string="postgresql://localhost/mydb")
shared_pool = await AsyncDatabaseManager.create_shared_pool(config)

# Create llmemory with shared pool
db_manager = AsyncDatabaseManager(pool=shared_pool, schema="llmemory")
memory = LLMemory.from_db_manager(
    db_manager,
    openai_api_key="sk-..."
)
await memory.initialize()

db_manager

Get the underlying database manager for health checks and monitoring.

Property:

@property
def db_manager(self) -> Optional[AsyncDatabaseManager]

Returns:

• Optional[AsyncDatabaseManager]: Database manager instance if initialized, None otherwise

Example:

from llmemory import LLMemory

memory = LLMemory(connection_string="postgresql://localhost/mydb")
await memory.initialize()

# Access underlying database manager
db_mgr = memory.db_manager
if db_mgr:
    # Check connection pool status
    pool_status = await db_mgr.get_pool_status()
    print(f"Active connections: {pool_status['active']}")
    print(f"Idle connections: {pool_status['idle']}")

    # Run health check
    is_healthy = await db_mgr.health_check()
    print(f"Database healthy: {is_healthy}")

When to use:

• Health monitoring and observability
• Accessing connection pool metrics
• Database diagnostics
• Integration with monitoring systems

initialize()

Initialize the library and database schema.

Signature:

async def initialize() -> None

Raises:

• DatabaseError: If database initialization fails
• ConfigurationError: If configuration is invalid

Example:

memory = LLMemory(connection_string="postgresql://localhost/mydb")
await memory.initialize()  # Sets up tables, migrations, indexes

close()

Close all connections and cleanup resources.

Signature:

async def close() -> None

Example:

await memory.close()

Context Manager Pattern (Recommended):

async with LLMemory(connection_string="...") as memory:
    # Use memory here
    results = await memory.search(...)
# Automatically closed

Document Types

class DocumentType(str, Enum):
    PDF = "pdf"
    MARKDOWN = "markdown"
    CODE = "code"
    TEXT = "text"
    HTML = "html"
    DOCX = "docx"
    EMAIL = "email"
    REPORT = "report"
    CHAT = "chat"
    PRESENTATION = "presentation"
    LEGAL_DOCUMENT = "legal_document"
    TECHNICAL_DOC = "technical_doc"
    BUSINESS_REPORT = "business_report"
    UNKNOWN = "unknown"

Search Types

class SearchType(str, Enum):
    VECTOR = "vector"     # Vector similarity search only
    TEXT = "text"         # Full-text search only
    HYBRID = "hybrid"     # Combines vector + text (recommended)

Chunking Strategies

class ChunkingStrategy(str, Enum):
    HIERARCHICAL = "hierarchical"      # Default - Creates parent and child chunks for better context
    FIXED_SIZE = "fixed_size"          # Fixed-size chunks with overlap
    SEMANTIC = "semantic"              # Chunks based on semantic boundaries (slower, higher quality)
    SLIDING_WINDOW = "sliding_window"  # Sliding window with configurable overlap

Strategy descriptions:

• HIERARCHICAL (default): Creates hierarchical parent and child chunks. Parent chunks provide broader context while child chunks are used for precise retrieval. Best for most use cases.
• FIXED_SIZE: Creates fixed-size chunks with configurable overlap. Simple and fast, good for uniform documents.
• SEMANTIC: Chunks based on semantic boundaries (paragraphs, sections). Slower but produces higher quality chunks that respect document structure.
• SLIDING_WINDOW: Creates overlapping chunks using a sliding window approach. Good for ensuring no information is lost at chunk boundaries.

Usage:

from llmemory import ChunkingStrategy

# Use enum value
result = await memory.add_document(
    owner_id="workspace-1",
    id_at_origin="user-123",
    document_name="example.txt",
    document_type=DocumentType.TEXT,
    content="Your document content...",
    chunking_strategy=ChunkingStrategy.SEMANTIC  # Use enum
)

# Or use string value (also valid)
result = await memory.add_document(
    owner_id="workspace-1",
    id_at_origin="user-123",
    document_name="example.txt",
    document_type=DocumentType.TEXT,
    content="Your document content...",
    chunking_strategy="hierarchical"  # String also works
)

Model Classes

SearchResult

Search result from any search operation.

Fields:

• chunk_id (UUID): Chunk identifier
• document_id (UUID): Document identifier
• content (str): Chunk content
• metadata (Dict[str, Any]): Chunk metadata
• score (float): Overall relevance score
• similarity (float, optional): Vector similarity score (0-1)
• text_rank (float, optional): Full-text search rank
• rrf_score (float, optional): Reciprocal Rank Fusion score
• rerank_score (float, optional): Reranker score (when reranking enabled)
• summary (str, optional): Chunk summary if generated
• parent_chunks (List[DocumentChunk]): Surrounding chunks if requested

EnrichedSearchResult

Extended search result with document metadata (inherits from SearchResult).

Additional Fields:

• document_name (str): Name of the source document
• document_type (str): Type of document
• document_metadata (Dict[str, Any]): Document-level metadata

When used: Returned by searchwithdocuments()

SearchResultWithDocuments

Container for enriched search results.

Fields:

• results (List[EnrichedSearchResult]): Enriched search results
• total (int): Total number of results

DocumentAddResult

Result of adding a document.

Fields:

• document (Document): Created document object with all fields
• chunks_created (int): Number of chunks created
• embeddings_created (int): Number of embeddings generated
• processingtimems (float): Processing time in milliseconds

DocumentListResult

Result of listing documents with pagination.

Fields:

• documents (List[Document]): Document objects
• total (int): Total matching documents (before pagination)
• limit (int): Applied limit
• offset (int): Applied offset

DocumentWithChunks

Document with optional chunks.

Fields:

• document (Document): Document object
• chunks (Optional[List[DocumentChunk]]): Chunks if requested
• chunk_count (int): Total number of chunks

OwnerStatistics

Statistics for an owner's documents.

Fields:

• document_count (int): Total documents
• chunk_count (int): Total chunks
• totalsizebytes (int): Estimated total size
• documenttypebreakdown (Optional[Dict[DocumentType, int]]): Count by document type
• createddaterange (Optional[Tuple[datetime, datetime]]): (mindate, maxdate) of document creation

DeleteResult

Result of batch delete operation.

Fields:

• deleted_count (int): Number of documents deleted
• deleteddocumentids (List[UUID]): IDs of deleted documents

EmbeddingStatus

Enum for embedding generation status.

class EmbeddingStatus(str, Enum):
    PENDING = "pending"        # Job queued but not started
    PROCESSING = "processing"  # Currently generating embeddings
    COMPLETED = "completed"    # Successfully completed
    FAILED = "failed"          # Failed with error

EmbeddingJob

Represents a background embedding generation job.

Fields:

• chunk_id (UUID): Chunk being processed
• provider_id (str): Embedding provider ID
• status (EmbeddingStatus): Current status
• retry_count (int): Number of retries attempted
• error_message (Optional[str]): Error details if failed
• created_at (datetime): When job was created
• processed_at (Optional[datetime]): When processing finished

SearchQuery

Internal search query model (rarely used directly).

Fields:

• owner_id (str): Owner identifier
• query_text (str): Search query text
• search_type (SearchType): Type of search
• limit (int): Maximum results
• alpha (float): Hybrid search weight
• metadata_filter (Optional[Dict[str, Any]]): Metadata filter
• idatorigin (Optional[str]): Single origin filter
• idatorigins (Optional[List[str]]): Multiple origins filter
• date_from (Optional[datetime]): Start date
• date_to (Optional[datetime]): End date
• includeparentcontext (bool): Include parent chunks
• context_window (int): Number of parent chunks
• rerank (bool): Enable reranking
• enablequeryexpansion (bool): Enable query expansion
• maxqueryvariants (int): Max query variants

add_document()

Add a document and process it into searchable chunks.

Signature:

async def add_document(
    owner_id: str,
    id_at_origin: str,
    document_name: str,
    document_type: Union[DocumentType, str],
    content: str,
    document_date: Optional[datetime] = None,
    metadata: Optional[Dict[str, Any]] = None,
    chunking_strategy: str = "hierarchical",
    chunking_config: Optional[ChunkingConfig] = None,
    generate_embeddings: bool = True
) -> DocumentAddResult

Parameters:

• owner_id (str, required): Owner identifier for multi-tenancy (e.g., "workspace-123", "tenant-abc")
• idatorigin (str, required): Origin identifier within owner (e.g., "user-456", "thread-789")
• document_name (str, required): Name of the document
• document_type (DocumentType or str, required): Type of document
• content (str, required): Full document content
• document_date (datetime, optional): Document date for temporal filtering
• metadata (Dict[str, Any], optional): Custom metadata (searchable via metadata_filter)
• chunking_strategy (str, default: "hierarchical"): Chunking strategy to use
• chunking_config (ChunkingConfig, optional): Custom chunking configuration
• generate_embeddings (bool, default: True): Generate embeddings immediately

Returns:

• DocumentAddResult with:

- document (Document): Created document object - chunkscreated (int): Number of chunks created - embeddingscreated (int): Number of embeddings generated - processingtimems (float): Processing time in milliseconds

Raises:

• ValidationError: If input validation fails (invalid owner_id, empty content, etc.)
• DatabaseError: If database operation fails
• EmbeddingError: If embedding generation fails

Example:

from llmemory import DocumentType
from datetime import datetime

result = await memory.add_document(
    owner_id="workspace-1",
    id_at_origin="user-123",
    document_name="Q4 Report.pdf",
    document_type=DocumentType.PDF,
    content="Full document text here...",
    document_date=datetime(2024, 10, 1),
    metadata={
        "category": "financial",
        "department": "finance",
        "confidential": False
    }
)

print(f"Document ID: {result.document.document_id}")
print(f"Chunks: {result.chunks_created}")
print(f"Embeddings: {result.embeddings_created}")
print(f"Time: {result.processing_time_ms:.2f}ms")

search()

Search for documents.

Signature:

async def search(
    owner_id: str,
    query_text: str,
    search_type: Union[SearchType, str] = SearchType.HYBRID,
    limit: int = 10,
    id_at_origin: Optional[str] = None,
    id_at_origins: Optional[List[str]] = None,
    metadata_filter: Optional[Dict[str, Any]] = None,
    date_from: Optional[datetime] = None,
    date_to: Optional[datetime] = None,
    include_parent_context: bool = False,
    context_window: int = 2,
    alpha: float = 0.5,
    query_expansion: Optional[bool] = None,
    max_query_variants: Optional[int] = None,
    rerank: Optional[bool] = None,
    rerank_top_k: Optional[int] = None,
    rerank_return_k: Optional[int] = None
) -> List[SearchResult]

Parameters:

• owner_id (str, required): Owner identifier for filtering
• query_text (str, required): Search query text
• search_type (SearchType or str, default: HYBRID): Type of search to perform
• limit (int, default: 10): Maximum number of results
• idatorigin (str, optional): Filter by single origin ID
• idatorigins (List[str], optional): Filter by multiple origin IDs
• metadata_filter (Dict[str, Any], optional): Filter by metadata (e.g., {"category": "financial"})
• date_from (datetime, optional): Start date filter
• date_to (datetime, optional): End date filter
• includeparentcontext (bool, default: False): Include surrounding chunks
• context_window (int, default: 2): Number of surrounding chunks to include
• alpha (float, default: 0.5): Hybrid search weight (0=text only, 1=vector only)
• query_expansion (bool, optional): Enable query expansion (None = follow config)
• maxqueryvariants (int, optional): Max query variants for expansion
• rerank (bool, optional): Enable reranking (None = follow config)
• reranktopk (int, optional): Candidates for reranker
• rerankreturnk (int, optional): Results after reranking

Returns:

• List[SearchResult] where each result has:

- chunkid (UUID): Chunk identifier - documentid (UUID): Document identifier - content (str): Chunk content - metadata (Dict[str, Any]): Chunk metadata - score (float): Overall relevance score - similarity (float, optional): Vector similarity score - textrank (float, optional): Text search rank - rrfscore (float, optional): Reciprocal Rank Fusion score - rerankscore (float, optional): Reranker score (when reranking enabled) - summary (str, optional): Chunk summary if available - parentchunks (List[DocumentChunk]): Surrounding chunks if requested

Raises:

• ValidationError: If input validation fails
• SearchError: If search operation fails

Example:

from llmemory import SearchType

# Basic search
results = await memory.search(
    owner_id="workspace-1",
    query_text="quarterly revenue trends",
    search_type=SearchType.HYBRID,
    limit=5
)

for result in results:
    print(f"Score: {result.score:.3f}")
    print(f"Content: {result.content[:100]}...")
    print(f"Metadata: {result.metadata}")
    print("---")

# Advanced search with filters
results = await memory.search(
    owner_id="workspace-1",
    query_text="product launch strategy",
    search_type=SearchType.HYBRID,
    limit=10,
    metadata_filter={"category": "strategy", "department": "product"},
    date_from=datetime(2024, 1, 1),
    date_to=datetime(2024, 12, 31),
    alpha=0.7  # Favor vector search slightly
)

searchwithdocuments()

Search and return results enriched with document metadata.

Signature:

async def search_with_documents(
    owner_id: str,
    query_text: str,
    search_type: Union[SearchType, str] = SearchType.HYBRID,
    limit: int = 10,
    metadata_filter: Optional[Dict[str, Any]] = None,
    include_document_metadata: bool = True
) -> SearchResultWithDocuments

Parameters:

• owner_id (str, required): Owner identifier
• query_text (str, required): Search query text
• search_type (SearchType or str, default: HYBRID): Type of search
• limit (int, default: 10): Maximum results
• metadata_filter (Dict[str, Any], optional): Filter by metadata
• includedocumentmetadata (bool, default: True): Include document-level metadata

Returns:

• SearchResultWithDocuments with:

- results (List[EnrichedSearchResult]): Enriched search results - total (int): Total number of results

EnrichedSearchResult fields:

• All fields from SearchResult (chunk_id, content, score, etc.)
• document_name (str): Name of the source document
• document_type (str): Type of document
• document_metadata (Dict[str, Any]): Document-level metadata

Raises:

• ValidationError: If input validation fails
• SearchError: If search operation fails

Example:

# Search with document context
results_with_docs = await memory.search_with_documents(
    owner_id="workspace-1",
    query_text="quarterly financial performance",
    search_type=SearchType.HYBRID,
    limit=10
)

print(f"Found {results_with_docs.total} results")

for result in results_with_docs.results:
    print(f"Document: {result.document_name}")
    print(f"Type: {result.document_type}")
    print(f"Score: {result.score:.3f}")
    print(f"Content: {result.content[:100]}...")
    print(f"Metadata: {result.document_metadata}")
    print("---")

When to use:

• When you need document context along with search results
• Building UI that shows source documents
• Grouping results by document
• When document metadata is needed for filtering or display

list_documents()

List documents with pagination and filtering.

Signature:

async def list_documents(
    owner_id: str,
    limit: int = 20,
    offset: int = 0,
    document_type: Optional[DocumentType] = None,
    order_by: Literal["created_at", "updated_at", "document_name"] = "created_at",
    order_desc: bool = True,
    metadata_filter: Optional[Dict[str, Any]] = None
) -> DocumentListResult

Parameters:

• owner_id (str, required): Owner identifier
• limit (int, default: 20): Maximum documents to return
• offset (int, default: 0): Number of documents to skip (for pagination)
• document_type (DocumentType, optional): Filter by document type
• orderby (str, default: "createdat"): Field to sort by
• order_desc (bool, default: True): Sort descending
• metadata_filter (Dict[str, Any], optional): Filter by metadata

Returns:

• DocumentListResult with:

- documents (List[Document]): Document objects - total (int): Total matching documents - limit (int): Applied limit - offset (int): Applied offset

Raises:

• ValidationError: If parameters are invalid

Example:

# List recent documents
result = await memory.list_documents(
    owner_id="workspace-1",
    limit=20,
    offset=0,
    order_by="created_at",
    order_desc=True
)

print(f"Total documents: {result.total}")
for doc in result.documents:
    print(f"{doc.document_name} - {doc.document_type.value}")

# Filter by type and metadata
result = await memory.list_documents(
    owner_id="workspace-1",
    document_type=DocumentType.PDF,
    metadata_filter={"category": "financial"},
    limit=50
)

get_document()

Retrieve a specific document with optional chunks.

Signature:

async def get_document(
    owner_id: str,
    document_id: Union[str, UUID],
    include_chunks: bool = False,
    include_embeddings: bool = False
) -> DocumentWithChunks

Parameters:

• owner_id (str, required): Owner/workspace identifier (required for access control)
• document_id (str or UUID, required): Document identifier
• include_chunks (bool, default: False): Include all chunks for this document
• includeembeddings (bool, default: False): Include embeddings with chunks (requires includechunks=True)

Returns:

• DocumentWithChunks with:

- document (Document): Document object - chunks (List[DocumentChunk], optional): Chunks if requested - chunk_count (int): Total number of chunks

Raises:

• DocumentNotFoundError: If document doesn't exist
• PermissionError: If the document belongs to a different owner

Example:

# Get document without chunks
doc_info = await memory.get_document(
    owner_id="workspace-1",
    document_id="uuid-here"
)
print(f"Document: {doc_info.document.document_name}")
print(f"Chunks: {doc_info.chunk_count}")

# Get document with all chunks
doc_with_chunks = await memory.get_document(
    owner_id="workspace-1",
    document_id="uuid-here",
    include_chunks=True
)

for chunk in doc_with_chunks.chunks:
    print(f"Chunk {chunk.chunk_index}: {chunk.content[:50]}...")

getdocumentchunks()

Get chunks for a specific document with pagination.

Signature:

async def get_document_chunks(
    owner_id: str,
    document_id: Union[str, UUID],
    limit: Optional[int] = None,
    offset: int = 0
) -> List[DocumentChunk]

Parameters:

• owner_id (str, required): Owner/workspace identifier (required for access control)
• document_id (str or UUID, required): Document identifier
• limit (int, optional): Maximum number of chunks to return (None = all chunks)
• offset (int, default: 0): Number of chunks to skip for pagination

Returns:

• List[DocumentChunk]: List of chunks ordered by chunk_index

Raises:

• DocumentNotFoundError: If document doesn't exist
• PermissionError: If the document belongs to a different owner
• ValidationError: If limit or offset are negative

Example:

# Get all chunks for a document
chunks = await memory.get_document_chunks(
    owner_id="workspace-1",
    document_id="uuid-here"
)
print(f"Total chunks: {len(chunks)}")
for chunk in chunks:
    print(f"Chunk {chunk.chunk_index}: {chunk.content[:50]}...")

# Paginated retrieval
page_size = 10
offset = 0
while True:
    chunks = await memory.get_document_chunks(
        owner_id="workspace-1",
        document_id="uuid-here",
        limit=page_size,
        offset=offset
    )

    if not chunks:
        break

    for chunk in chunks:
        print(f"Chunk {chunk.chunk_index}: {chunk.content}")

    offset += page_size

When to use:

• Accessing document chunks without full document
• Paginating through large documents
• Processing chunks in batches
• Inspecting chunking results

getchunkcount()

Get the number of chunks for a document.

Signature:

async def get_chunk_count(
    owner_id: str,
    document_id: Union[str, UUID]
) -> int

Parameters:

• owner_id (str, required): Owner/workspace identifier (required for access control)
• document_id (str or UUID, required): Document identifier

Returns:

• int: Number of chunks for the document

Raises:

• DocumentNotFoundError: If document doesn't exist
• PermissionError: If the document belongs to a different owner

Example:

# Get chunk count
count = await memory.get_chunk_count(owner_id="workspace-1", document_id="uuid-here")
print(f"Document has {count} chunks")

# Check if document needs re-chunking
if count > 1000:
    print("Warning: Very large document, consider splitting")
elif count == 0:
    print("Warning: Document has no chunks")

When to use:

• Quick check of document size
• Validating chunking results
• Deciding pagination strategy
• Monitoring document processing

delete_document()

Delete a single document and all its chunks.

Signature:

async def delete_document(
    owner_id: str,
    document_id: Union[UUID, str]
) -> None

Parameters:

• owner_id (str, required): Owner/workspace identifier (required for access control)
• document_id (UUID or str, required): Document ID to delete

Raises:

• ResourceNotFoundError: If document not found
• PermissionError: If the document belongs to a different owner
• DatabaseError: If deletion fails

Example:

await memory.delete_document(owner_id="workspace-1", document_id="uuid-here")

delete_documents()

Delete multiple documents.

Signature:

async def delete_documents(
    owner_id: str,
    document_ids: Optional[List[Union[str, UUID]]] = None,
    metadata_filter: Optional[Dict[str, Any]] = None
) -> DeleteResult

Parameters:

• owner_id (str, required): Owner identifier (safety check)
• document_ids (List[UUID or str], optional): Specific documents to delete
• metadata_filter (Dict[str, Any], optional): Delete all matching metadata

Returns:

• DeleteResult with:

- deletedcount (int): Number of documents deleted - deleteddocument_ids (List[UUID]): IDs of deleted documents

Raises:

• ValueError: If neither documentids nor metadatafilter provided
• ValidationError: If owner_id is invalid

Example:

# Delete specific documents
result = await memory.delete_documents(
    owner_id="workspace-1",
    document_ids=["uuid-1", "uuid-2", "uuid-3"]
)
print(f"Deleted {result.deleted_count} documents")

# Delete by metadata
result = await memory.delete_documents(
    owner_id="workspace-1",
    metadata_filter={"category": "temp", "delete_after": "2024-01-01"}
)

get_statistics()

Get statistics for an owner's documents.

Signature:

async def get_statistics(
    owner_id: str,
    include_breakdown: bool = False
) -> OwnerStatistics

Parameters:

• owner_id (str, required): Owner identifier
• include_breakdown (bool, default: False): Include breakdown by document type

Returns:

• OwnerStatistics with:

- documentcount (int): Total documents - chunkcount (int): Total chunks - totalsizebytes (int): Estimated total size - documenttypebreakdown (Dict[DocumentType, int], optional): Count by type - createddaterange (Tuple[datetime, datetime], optional): Date range

Example:

stats = await memory.get_statistics(
    owner_id="workspace-1",
    include_breakdown=True
)

print(f"Documents: {stats.document_count}")
print(f"Chunks: {stats.chunk_count}")
print(f"Size: {stats.total_size_bytes / 1024 / 1024:.2f} MB")

if stats.document_type_breakdown:
    for doc_type, count in stats.document_type_breakdown.items():
        print(f"  {doc_type.value}: {count}")

Common Patterns

Async Context Manager (Recommended)

async with LLMemory(connection_string="postgresql://localhost/mydb") as memory:
    # Add documents
    await memory.add_document(...)

    # Search
    results = await memory.search(...)
# Automatically closed

Batch Document Processing

documents = [
    {"name": "doc1.txt", "content": "..."},
    {"name": "doc2.txt", "content": "..."},
    {"name": "doc3.txt", "content": "..."},
]

for doc in documents:
    result = await memory.add_document(
        owner_id="workspace-1",
        id_at_origin="batch-import",
        document_name=doc["name"],
        document_type=DocumentType.TEXT,
        content=doc["content"]
    )
    print(f"Added {doc['name']}: {result.chunks_created} chunks")

Filtered Search with Metadata

# Add document with metadata
await memory.add_document(
    owner_id="workspace-1",
    id_at_origin="user-123",
    document_name="report.pdf",
    document_type=DocumentType.PDF,
    content="...",
    metadata={
        "category": "financial",
        "year": 2024,
        "quarter": "Q4",
        "confidential": False
    }
)

# Search with metadata filter
results = await memory.search(
    owner_id="workspace-1",
    query_text="revenue analysis",
    metadata_filter={
        "category": "financial",
        "year": 2024
    },
    limit=10
)

Paginated Document Listing

page_size = 20
offset = 0

while True:
    result = await memory.list_documents(
        owner_id="workspace-1",
        limit=page_size,
        offset=offset
    )

    if not result.documents:
        break

    for doc in result.documents:
        print(f"{doc.document_name}: {doc.chunk_count} chunks")

    offset += page_size
    if offset >= result.total:
        break

Exception Reference

All llmemory exceptions inherit from LLMemoryError base class.

Exception Hierarchy

LLMemoryError (base)
├── ConfigurationError
├── ValidationError
├── DatabaseError
│   └── ConnectionError
├── EmbeddingError
├── SearchError
├── ChunkingError
├── ResourceNotFoundError
│   └── DocumentNotFoundError
├── RateLimitError
└── PermissionError

LLMemoryError

Base exception for all llmemory errors.

When raised: Never raised directly, use specific subclasses

Usage:

from llmemory import LLMemoryError

try:
    await memory.search(...)
except LLMemoryError as e:
    # Catches all llmemory exceptions
    print(f"LLMemory error: {e}")

ConfigurationError

Configuration is invalid or incomplete.

Common causes:

• Missing required configuration (connection_string, API key)
• Invalid configuration values (negative pool size, invalid dimensions)
• Incompatible configuration combinations

When raised:

• During LLMemory() initialization if neither connectionstring nor dbmanager provided
• During initialize() if config validation fails
• When embedding provider configuration is invalid

Example:

from llmemory import ConfigurationError

try:
    # Missing connection_string
    memory = LLMemory()  # Raises ConfigurationError
except ConfigurationError as e:
    print(f"Invalid configuration: {e}")

ValidationError

Input validation failed.

Common causes:

• owner_id too long or invalid characters
• Empty or too long content
• Invalid document_name
• Negative limit or offset values

When raised:

• During adddocument() if ownerid, idatorigin, or content invalid
• During search() if ownerid or querytext invalid
• During list_documents() if pagination parameters invalid

Example:

from llmemory import ValidationError

try:
    await memory.add_document(
        owner_id="",  # Empty owner_id - invalid
        id_at_origin="user-123",
        document_name="doc.txt",
        document_type=DocumentType.TEXT,
        content="content"
    )
except ValidationError as e:
    print(f"Validation failed: {e}")
    # Output: "Validation failed: owner_id cannot be empty"

DatabaseError

Database operation failed.

Common causes:

• Connection to PostgreSQL failed
• Query execution failed
• Transaction rollback
• Schema migration failed

When raised:

• During initialize() if database setup fails
• During any CRUD operation if database query fails
• During add_document() if insert fails

Example:

from llmemory import DatabaseError

try:
    await memory.add_document(...)
except DatabaseError as e:
    print(f"Database error: {e}")
    # Possible causes: connection lost, disk full, constraint violation

ConnectionError

Cannot connect to database (subclass of DatabaseError).

Common causes:

• PostgreSQL not running
• Wrong connection string
• Network issues
• Firewall blocking connection

When raised:

• During initialize() if connection fails
• During operations if connection is lost

Example:

from llmemory import ConnectionError

try:
    memory = LLMemory(connection_string="postgresql://invalid:5432/db")
    await memory.initialize()
except ConnectionError as e:
    print(f"Cannot connect to database: {e}")

EmbeddingError

Embedding generation failed.

Common causes:

• OpenAI API key invalid or missing
• OpenAI rate limit exceeded
• Local embedding model failed to load
• Invalid embedding dimensions

When raised:

• During adddocument() if generateembeddings=True and embedding fails
• During processpendingembeddings() if batch processing fails

Example:

from llmemory import EmbeddingError

try:
    await memory.add_document(
        owner_id="workspace-1",
        id_at_origin="user-123",
        document_name="doc.txt",
        document_type=DocumentType.TEXT,
        content="content",
        generate_embeddings=True  # Will fail if no API key
    )
except EmbeddingError as e:
    print(f"Embedding generation failed: {e}")

SearchError

Search operation failed.

Common causes:

• Invalid search query syntax
• Vector index not built
• Embedding provider not configured for vector search
• Search timeout exceeded

When raised:

• During search() if query execution fails
• During vector search if embeddings table doesn't exist
• During hybrid search if either vector or text search fails

Example:

from llmemory import SearchError

try:
    results = await memory.search(
        owner_id="workspace-1",
        query_text="test",
        search_type=SearchType.VECTOR  # Fails if no embeddings
    )
except SearchError as e:
    print(f"Search failed: {e}")

ChunkingError

Document chunking failed.

Common causes:

• Invalid chunking configuration
• Document too large to chunk
• Chunking strategy not supported for document type

When raised:

• During add_document() if chunking fails
• During process_document() if chunker fails

Example:

from llmemory import ChunkingError

try:
    await memory.add_document(
        owner_id="workspace-1",
        id_at_origin="user-123",
        document_name="huge.txt",
        document_type=DocumentType.TEXT,
        content="x" * 100_000_000  # Too large
    )
except ChunkingError as e:
    print(f"Chunking failed: {e}")

ResourceNotFoundError

Requested resource doesn't exist.

Common causes:

• Document ID doesn't exist
• Chunk ID not found
• Owner has no documents

When raised:

• During delete_document() if document not found
• During get_document() if document doesn't exist

DocumentNotFoundError

Specific document doesn't exist (subclass of ResourceNotFoundError).

When raised:

• During getdocument() if documentid doesn't exist
• During delete_document() if document not found

Example:

from llmemory import DocumentNotFoundError
from uuid import UUID

try:
    doc = await memory.get_document(
        owner_id="workspace-1",
        document_id=UUID("00000000-0000-0000-0000-000000000000")
    )
except DocumentNotFoundError as e:
    print(f"Document not found: {e}")

RateLimitError

API rate limit exceeded.

Common causes:

• OpenAI API rate limit hit
• Too many embedding requests in short time
• Exceeded configured rate limits

When raised:

• During embedding generation if API rate limited
• During query expansion if LLM API rate limited

Example:

from llmemory import RateLimitError
import asyncio

try:
    # Batch process with rate limiting
    for doc in documents:
        await memory.add_document(...)
except RateLimitError as e:
    print(f"Rate limited: {e}")
    await asyncio.sleep(60)  # Wait before retry

PermissionError

Permission denied for operation.

Common causes:

• Attempting to access document owned by different owner_id
• Database permission denied

When raised:

• During operations if user doesn't have permission
• During delete if document belongs to different owner

Example:

from llmemory import PermissionError as LLMemoryPermissionError

try:
    # Trying to access another owner's document
    doc = await memory.get_document(owner_id="workspace-1", document_id="...")
except LLMemoryPermissionError as e:
    print(f"Permission denied: {e}")

Error Handling Patterns

Basic Error Handling

from llmemory import (
    LLMemoryError, ConfigurationError, ValidationError, DatabaseError,
    DocumentNotFoundError, EmbeddingError, SearchError, ChunkingError,
    ResourceNotFoundError, RateLimitError, ConnectionError
)

try:
    memory = LLMemory(connection_string="postgresql://localhost/mydb")
    await memory.initialize()

    result = await memory.add_document(
        owner_id="workspace-1",
        id_at_origin="user-123",
        document_name="test.txt",
        document_type=DocumentType.TEXT,
        content="Test content"
    )

    results = await memory.search(
        owner_id="workspace-1",
        query_text="test query"
    )

except ConfigurationError as e:
    print(f"Configuration error: {e}")
except ValidationError as e:
    print(f"Validation error: {e}")
except ConnectionError as e:
    print(f"Cannot connect to database: {e}")
except DatabaseError as e:
    print(f"Database error: {e}")
except DocumentNotFoundError as e:
    print(f"Document not found: {e}")
except EmbeddingError as e:
    print(f"Embedding error: {e}")
except SearchError as e:
    print(f"Search error: {e}")
except ChunkingError as e:
    print(f"Chunking error: {e}")
except RateLimitError as e:
    print(f"Rate limit hit: {e}")
    await asyncio.sleep(60)  # Wait before retry
except LLMemoryError as e:
    print(f"Unexpected llmemory error: {e}")
finally:
    await memory.close()

Granular Error Handling

# Handle specific errors differently
try:
    result = await memory.add_document(...)
except ValidationError as e:
    # User input error - return 400
    return {"error": str(e), "code": 400}
except EmbeddingError as e:
    # Embedding failed but document added - return partial success
    logger.error(f"Embedding failed: {e}")
    return {"warning": "Document added but embeddings pending", "code": 202}
except DatabaseError as e:
    # System error - return 500
    logger.error(f"Database error: {e}")
    return {"error": "Internal server error", "code": 500}

Retry Logic for Transient Errors

import asyncio
from llmemory import RateLimitError, ConnectionError

async def robust_search(memory, owner_id, query, max_retries=3):
    """Search with retry logic for transient errors."""
    for attempt in range(max_retries):
        try:
            return await memory.search(
                owner_id=owner_id,
                query_text=query
            )
        except RateLimitError:
            if attempt < max_retries - 1:
                await asyncio.sleep(2 ** attempt)  # Exponential backoff
                continue
            raise
        except ConnectionError:
            if attempt < max_retries - 1:
                await asyncio.sleep(1)
                continue
            raise

Complete Environment Variable Reference

Database Configuration

DATABASE_URL=postgresql://localhost/mydb  # PostgreSQL connection string
LLMEMORY_DB_MIN_POOL_SIZE=5              # Minimum connection pool size (default: 5)
LLMEMORY_DB_MAX_POOL_SIZE=20             # Maximum connection pool size (default: 20)

Embedding Configuration

# Provider selection
OPENAI_API_KEY=sk-...                    # OpenAI API key (required for OpenAI embeddings)
LLMEMORY_EMBEDDING_PROVIDER=openai       # Provider: "openai" or "local-minilm" (default: "openai")

# Local embedding models
LLMEMORY_LOCAL_MODEL=all-MiniLM-L6-v2    # Local model name (default: all-MiniLM-L6-v2)
LLMEMORY_LOCAL_DEVICE=cpu                # Device: "cpu" or "cuda" (default: cpu)
LLMEMORY_LOCAL_CACHE_DIR=/path/to/cache  # Cache directory for local models

Search Configuration

# HNSW Index tuning
LLMEMORY_HNSW_PROFILE=balanced           # Profile: "fast", "balanced", "accurate" (default: balanced)

# Search defaults
LLMEMORY_DEFAULT_SEARCH_TYPE=hybrid      # Default search type (default: hybrid)
LLMEMORY_SEARCH_CACHE_TTL=300            # Search cache TTL in seconds (default: 300)

Query Expansion Configuration

LLMEMORY_ENABLE_QUERY_EXPANSION=1        # Enable query expansion: 1 or 0 (default: 0)
LLMEMORY_MAX_QUERY_VARIANTS=3            # Max query variants to generate (default: 3)

Reranking Configuration

LLMEMORY_ENABLE_RERANK=1                 # Enable reranking: 1 or 0 (default: 0)
LLMEMORY_RERANK_PROVIDER=openai          # Provider: "openai", "lexical" (default: lexical)
LLMEMORY_RERANK_MODEL=gpt-4.1-mini       # Reranking model name
LLMEMORY_RERANK_TOP_K=50                 # Candidates to consider (default: 50)
LLMEMORY_RERANK_RETURN_K=15              # Results to return after reranking (default: 15)
LLMEMORY_RERANK_DEVICE=cpu               # Device for local rerankers: "cpu" or "cuda"
LLMEMORY_RERANK_BATCH_SIZE=16            # Batch size for local reranking (default: 16)

Chunking Configuration

LLMEMORY_ENABLE_CHUNK_SUMMARIES=1        # Enable chunk summaries: 1 or 0 (default: 0)

Feature Flags

LLMEMORY_DISABLE_CACHING=1               # Disable search caching (default: enabled)
LLMEMORY_DISABLE_METRICS=1               # Disable Prometheus metrics (default: enabled)

Logging

LLMEMORY_LOG_LEVEL=INFO                  # Log level: DEBUG, INFO, WARNING, ERROR (default: INFO)

Complete Configuration Reference

LLMemoryConfig

Main configuration class containing all subsystem configurations.

Constructor:

LLMemoryConfig(
    embedding: EmbeddingConfig = EmbeddingConfig(),
    chunking: ChunkingConfig = ChunkingConfig(),
    search: SearchConfig = SearchConfig(),
    database: DatabaseConfig = DatabaseConfig(),
    validation: ValidationConfig = ValidationConfig(),
    enable_caching: bool = True,
    enable_metrics: bool = True,
    enable_background_processing: bool = True,
    log_level: str = "INFO",
    log_slow_queries: bool = True,
    slow_query_threshold: float = 1.0
)

Creating and using config:

from llmemory import LLMemoryConfig

# Use default configuration
config = LLMemoryConfig()

# Modify specific settings
config.embedding.default_provider = "openai"
config.chunking.default_parent_size = 1000
config.search.enable_query_expansion = True

# Use with LLMemory
memory = LLMemory(
    connection_string="postgresql://localhost/mydb",
    config=config
)

Loading from environment:

# Automatically reads from environment variables
config = LLMemoryConfig.from_env()
memory = LLMemory(connection_string="...", config=config)

EmbeddingConfig

Configuration for embedding generation.

Fields:

• default_provider (str, default: "openai"): Default embedding provider
• providers (Dict[str, EmbeddingProviderConfig]): Available providers
• autocreatetables (bool, default: True): Auto-create provider tables

Example:

config = LLMemoryConfig()
config.embedding.default_provider = "local-minilm"

EmbeddingProviderConfig

Configuration for a single embedding provider.

Fields:

• provider_type (str): "openai" or "local"
• model_name (str): Model name
• dimension (int): Embedding dimensions
• api_key (Optional[str]): API key (for OpenAI)
• device (str, default: "cpu"): Device for local models ("cpu" or "cuda")
• cache_dir (Optional[str]): Cache directory for local models
• batch_size (int, default: 100): Batch size for processing
• max_retries (int, default: 3): Max retries on failure
• retry_delay (float, default: 1.0): Delay between retries in seconds
• timeout (float, default: 30.0): Request timeout in seconds
• maxtokensper_minute (int, default: 1,000,000): Rate limit for tokens
• maxrequestsper_minute (int, default: 3,000): Rate limit for requests

ChunkingConfig

Configuration for document chunking (in config.py).

Fields:

• enablechunksummaries (bool, default: False): Generate summaries for chunks
• summarymaxtokens (int, default: 120): Max tokens for summaries
• minchunksize (int, default: 50): Minimum chunk size in tokens
• maxchunksize (int, default: 2000): Maximum chunk size in tokens
• enablecontextualretrieval (bool, default: False): Prepend document context to chunks before embedding (Anthropic's approach)
• contexttemplate (str): Template for contextual retrieval format (default: "Document: {documentname}\nType: {document_type}\n\n{content}")

Contextual Retrieval Example:

config = LLMemoryConfig()
config.chunking.enable_contextual_retrieval = True

memory = LLMemory(connection_string="...", config=config)

# Chunks are embedded with document context prepended:
# "Document: Q3 Report\nType: report\n\nRevenue increased 15%"
#
# But chunk.content remains original for display:
# "Revenue increased 15%"

await memory.add_document(
    owner_id="workspace-1",
    id_at_origin="kb",
    document_name="Q3 Report",
    document_type=DocumentType.REPORT,
    content="Revenue increased 15% QoQ..."
)

Example:

config = LLMemoryConfig()
config.chunking.enable_chunk_summaries = True
config.chunking.summary_max_tokens = 100

SearchConfig

Configuration for search operations.

Fields:

• default_limit (int, default: 10): Default result limit
• max_limit (int, default: 100): Maximum allowed limit
• defaultsearchtype (str, default: "hybrid"): Default search type
• hnsw_profile (str, default: "balanced"): HNSW index profile
• rrf_k (int, default: 50): RRF constant for fusion
• enablequeryexpansion (bool, default: False): Enable query expansion
• maxqueryvariants (int, default: 3): Max query variants
• queryexpansionmodel (Optional[str]): Model for expansion
• includekeywordvariant (bool, default: True): Include keyword variant
• enable_rerank (bool, default: False): Enable reranking
• defaultrerankmodel (Optional[str]): Reranking model
• rerank_provider (str, default: "lexical"): Reranker provider
• reranktopk (int, default: 50): Candidates for reranking
• rerankreturnk (int, default: 15): Results after reranking
• rerank_device (Optional[str]): Device for local rerankers
• rerankbatchsize (int, default: 16): Batch size for reranking
• hnswefsearch (int, default: 100): HNSW ef_search parameter
• vectorsearchlimit (int, default: 100): Internal vector search limit
• textsearchlimit (int, default: 100): Internal text search limit
• cache_ttl (int, default: 3600): Cache TTL in seconds
• cachemaxsize (int, default: 10000): Max cache entries
• search_timeout (float, default: 5.0): Search timeout in seconds
• minscorethreshold (float, default: 0.0): Minimum score threshold

Example:

config = LLMemoryConfig()
config.search.enable_query_expansion = True
config.search.enable_rerank = True
config.search.rerank_provider = "openai"
config.search.hnsw_profile = "accurate"

DatabaseConfig

Configuration for database operations.

Fields:

• minpoolsize (int, default: 5): Minimum connection pool size
• maxpoolsize (int, default: 20): Maximum connection pool size
• connection_timeout (float, default: 10.0): Connection timeout in seconds
• command_timeout (float, default: 30.0): Command timeout in seconds
• schema_name (str, default: "llmemory"): PostgreSQL schema name
• documents_table (str, default: "documents"): Documents table name
• chunkstable (str, default: "documentchunks"): Chunks table name
• embeddingsqueuetable (str, default: "embedding_queue"): Queue table name
• searchhistorytable (str, default: "search_history"): Search history table
• embeddingproviderstable (str, default: "embedding_providers"): Providers table
• chunkembeddingsprefix (str, default: "chunkembeddings"): Embedding table prefix
• hnswindexname (str, default: "documentchunksembedding_hnsw"): HNSW index name
• hnsw_m (int, default: 16): HNSW M parameter
• hnswefconstruction (int, default: 200): HNSW ef_construction parameter

Example:

config = LLMemoryConfig()
config.database.schema_name = "my_app_llmemory"
config.database.min_pool_size = 10
config.database.max_pool_size = 50

ValidationConfig

Configuration for input validation.

Fields:

• maxowneridlength (int, default: 255): Max ownerid length
• maxidatoriginlength (int, default: 255): Max idatorigin length
• maxdocumentname_length (int, default: 500): Max document name length
• maxcontentlength (int, default: 10,000,000): Max content length (10MB)
• maxmetadatasize (int, default: 65536): Max metadata size (64KB)
• mincontentlength (int, default: 10): Minimum content length
• validowneridpattern (str): Regex for valid ownerid
• valididatoriginpattern (str): Regex for valid idatorigin

Example:

config = LLMemoryConfig()
config.validation.max_content_length = 20_000_000  # 20MB
config.validation.min_content_length = 50  # Require at least 50 chars

Common Mistakes

❌ Wrong: Not calling initialize()

memory = LLMemory(connection_string="...")
results = await memory.search(...)  # Error: not initialized

✅ Right: Always call initialize()

memory = LLMemory(connection_string="...")
await memory.initialize()  # Required!
results = await memory.search(...)

❌ Wrong: Not closing connections

memory = LLMemory(connection_string="...")
await memory.initialize()
# ... use memory ...
# Missing: await memory.close()

✅ Right: Use context manager

async with LLMemory(connection_string="...") as memory:
    # ... use memory ...
# Automatically closed

❌ Wrong: Forgetting owner_id filtering

results = await memory.search(
    owner_id="workspace-1",
    query_text="sensitive data"
)
# Results only from workspace-1 (good!)
# But need to verify owner_id matches current user

✅ Right: Always validate owner_id

current_workspace = get_current_workspace()
results = await memory.search(
    owner_id=current_workspace,  # Validated owner
    query_text="sensitive data"
)

Related Skills

• hybrid-search - Vector + BM25 hybrid search patterns
• multi-query - Query expansion and multi-query retrieval
• multi-tenant - Multi-tenant isolation patterns for SaaS
• rag - Building complete RAG systems with reranking

Important Notes

Multi-Tenancy: Always provide owner_id for proper data isolation. llmemory automatically filters all operations by owner.

Connection Pooling: For production applications with multiple services, use fromdbmanager() with a shared connection pool (see pgdbm-shared-pool skill).

Chunking: Documents are automatically chunked during add_document(). Default strategy is hierarchical chunking which creates parent and child chunks for better retrieval.

Embeddings: Embeddings are generated automatically unless generate_embeddings=False. For batch operations, consider using background processing.

Search Types:

• VECTOR: Best for semantic similarity
• TEXT: Best for exact keyword matching
• HYBRID: Best for most use cases (combines both)