v0.6.2 — Retrieval Strategies, Memory & Tools

What's New

Retrieval (4 new strategies)

• CRAGRetriever — Corrective RAG: grades retrieved docs for relevance, rewrites query and retries when too few are relevant
• QueryDecompositionRetriever — breaks complex queries into sub-queries, retrieves for each, deduplicates
• ContextualCompressionRetriever — compresses retrieved documents to only relevant excerpts using an LLM
• EnsembleRetriever — fuses results from multiple retrievers using weighted Reciprocal Rank Fusion

Memory (2 new backends)

• SQLiteConversationMemory — persistent chat history in SQLite, multi-conversation support, optional sliding window
• SummaryBufferMemory — token-budget-aware memory that progressively summarizes older messages

Tools (2 new)

• HumanInputTool — pauses agent execution to ask the user a question (custom sync/async input functions)
• WikipediaTool — searches and fetches Wikipedia article summaries via REST API, no extra deps

Stats

• 512 tests passing
• 13 LLM providers
• 13 built-in tools
• 12 document loaders
• 10 retrieval strategies
• 4 memory backends

pip install synapsekit==0.6.2