v1.4.2 — HuggingFace, Cache Backends, Graph Versioning

What's new in v1.4.2

New LLM Provider

• HuggingFaceLLM — Hugging Face Inference API via AsyncInferenceClient; supports both serverless API (model ID) and Dedicated Inference Endpoints (URL); streaming and generate

New Cache Backends

• DynamoDBCacheBackend — serverless LLM response caching on AWS DynamoDB; configurable partition key, TTL via ttl_seconds, custom region/endpoint (pip install synapsekit[dynamodb])
• MemcachedCacheBackend — high-throughput distributed LLM caching via aiomcache; TTL via exptime (pip install synapsekit[memcached])

New Tools

• GoogleSearchTool — Google web search via SerpAPI; graceful error handling for missing keys, empty results, API errors (pip install synapsekit[google-search])

Graph Improvements

• Graph versioning + checkpoint migration — StateGraph(version="1", migrations={"1": fn, ...}); CompiledGraph.resume() applies migration chains when loading older checkpoint versions; supports direct and chained (next_version, state_dict) migrations; missing paths raise GraphRuntimeError

Improved

• SQLQueryTool — parameterized queries via params dict (eliminates string interpolation risk); max_rows cap; fixed variable name shadowing

---

Stats: 1368 tests · 26 LLM providers · 41 tools · 15 loaders · 6 cache backends

---

Thanks to @DhruvGarg111 for all 6 contributions in this release!