vendor_connectors.meshy.persistence.vector_store

vendor_connectors.meshy.persistence.vector_store

SQLite vector store for idempotency tracking and RAG embeddings.

This module provides a vector-enabled SQLite database for:

1. Idempotency tracking - deduplicate asset generation requests
2. RAG embeddings - semantic search over prompts, species, assets

Uses sqlite-vec for vector similarity search, falling back to standard SQLite if vectors aren’t needed.

Usage: from vendor_connectors.meshy.persistence.vector_store import VectorStore

store = VectorStore("assets.db")

# Track asset generation
store.record_generation(
    spec_hash="abc123",
    prompt="cute otter character",
    task_id="meshy-task-id",
    embedding=get_embedding("cute otter character"),
)

# Find similar prompts (RAG)
similar = store.search_similar("river otter", limit=5)

# Check idempotency
existing = store.get_by_spec_hash("abc123")

Requirements: pip install mesh-toolkit[vector]

The vector extra includes:
- sqlite-vec (vector similarity extension)
- Optional: sentence-transformers for embeddings

Module Contents

Classes

GenerationRecord	Record of a 3D asset generation.
SimilarityResult	Result from similarity search.
VectorStore	SQLite vector store for asset generation tracking and RAG.

Functions

_utc_now
get_embedding	Get embedding for text using sentence-transformers.

Data

_HAS_VECTOR

API

vendor_connectors.meshy.persistence.vector_store._HAS_VECTOR

False

vendor_connectors.meshy.persistence.vector_store._utc_now() → datetime.datetime

class vendor_connectors.meshy.persistence.vector_store.GenerationRecord

Record of a 3D asset generation.

id : int | None

None

spec_hash : str =

project : str =

prompt : str =

art_style : str

‘sculpture’

task_id : str | None

None

status : str

‘pending’

model_url : str | None

None

embedding : list[float] | None

None

metadata : dict[str, Any]

‘field(…)’

created_at : datetime.datetime

‘field(…)’

updated_at : datetime.datetime

‘field(…)’

class vendor_connectors.meshy.persistence.vector_store.SimilarityResult

Result from similarity search.

record : vendor_connectors.meshy.persistence.vector_store.GenerationRecord

None

distance : float

None

score : float

None

class vendor_connectors.meshy.persistence.vector_store.VectorStore(db_path: str | pathlib.Path = ‘vendor_connectors.meshy.db’, embedding_dim: int = DEFAULT_EMBEDDING_DIM)

SQLite vector store for asset generation tracking and RAG.

Features:

• Idempotent task tracking by spec_hash
• Vector similarity search for RAG
• Full-text search fallback when vectors unavailable
• Atomic transactions

Args: db_path: Path to SQLite database file embedding_dim: Dimension of embedding vectors (default: 384 for MiniLM)

Initialization

DEFAULT_EMBEDDING_DIM

384

_get_conn() → sqlite3.Connection

Get or create database connection.

_transaction() → collections.abc.Iterator[sqlite3.Connection]

Context manager for transactions.

_init_db() → None

Initialize database schema.

record_generation(spec_hash: str, prompt: str, project: str = ‘default’, art_style: str = ‘sculpture’, task_id: str | None = None, embedding: list[float] | None = None, metadata: dict[str, Any] | None = None) → vendor_connectors.meshy.persistence.vector_store.GenerationRecord

Record a new generation (idempotent by spec_hash).

If a record with the same spec_hash exists, returns existing.

Args: spec_hash: Unique hash of the generation spec prompt: Text prompt for generation project: Project identifier art_style: Art style (realistic, sculpture, etc.) task_id: Meshy task ID if already submitted embedding: Optional embedding vector for RAG metadata: Additional metadata dict

Returns: GenerationRecord (existing or newly created)

update_status(spec_hash: str, status: str, task_id: str | None = None, model_url: str | None = None) → bool

Update generation status.

Args: spec_hash: Generation spec hash status: New status task_id: Optional task ID to update model_url: Optional model URL to update

Returns: True if record was updated, False if not found

get_by_spec_hash(spec_hash: str) → vendor_connectors.meshy.persistence.vector_store.GenerationRecord | None

Get generation record by spec hash.

Args: spec_hash: Generation spec hash

Returns: GenerationRecord or None

get_by_task_id(task_id: str) → vendor_connectors.meshy.persistence.vector_store.GenerationRecord | None

Get generation record by Meshy task ID.

Args: task_id: Meshy task ID

Returns: GenerationRecord or None

search_similar(query_embedding: list[float], limit: int = 10, project: str | None = None) → list[vendor_connectors.meshy.persistence.vector_store.SimilarityResult]

Search for similar generations using vector similarity.

Args: query_embedding: Query embedding vector limit: Maximum results to return project: Optional project filter

Returns: List of SimilarityResult ordered by similarity (highest first)

search_text(query: str, limit: int = 10, project: str | None = None) → list[vendor_connectors.meshy.persistence.vector_store.GenerationRecord]

Full-text search for prompts.

Falls back to this when vector search is unavailable.

Args: query: Search query limit: Maximum results project: Optional project filter

Returns: List of matching GenerationRecords

list_pending(project: str | None = None) → list[vendor_connectors.meshy.persistence.vector_store.GenerationRecord]

List all pending/in-progress generations.

Args: project: Optional project filter

Returns: List of pending GenerationRecords

compute_spec_hash(spec: dict[str, Any]) → str

Compute deterministic hash for a generation spec.

Args: spec: Generation specification dict

Returns: SHA256 hex digest

_row_to_record(row: sqlite3.Row) → vendor_connectors.meshy.persistence.vector_store.GenerationRecord

Convert database row to GenerationRecord.

_serialize_embedding(embedding: list[float]) → bytes

Serialize embedding to bytes for SQLite vec.

close() → None

Close database connection.

vendor_connectors.meshy.persistence.vector_store.get_embedding(text: str, model: str = ‘all-MiniLM-L6-v2’) → list[float] | None

Get embedding for text using sentence-transformers.

Args: text: Text to embed model: Model name (default: all-MiniLM-L6-v2)

Returns: Embedding vector or None if sentence-transformers not available