akashi_rag

Releases

1. ▸▾0.1.030 Jun 2026

   Initial release of retrieval-augmented generation for Akashi.

   • Initial release of retrieval-augmented generation for Akashi.
   • Retriever — the single read-side contract agents consume; the built-in KnowledgeBase and any external/"standard" RAG service satisfy it the same way.
   • Document / Chunk / EmbeddedChunk value types, plus RetrievalQuery / RetrievedChunk.
   • Chunker with RecursiveChunker (boundary-aware, the default) and FixedSizeChunker (sliding window). Character-based sizing with overlap.
   • VectorStore contract and InMemoryVectorStore — a pure-Dart, brute-force cosine-similarity index with metadata filtering and toJson / fromJson persistence. Runs offline with no dependencies.
   • KnowledgeBase — the built-in façade pairing a core EmbeddingModel with a VectorStore: chunk → embed (batched) → upsert on ingest, embed → search on retrieve. Implements Retriever.
   • retrievalTool — wraps a Retriever as an Akashi Tool (model-driven retrieval), and renderChunks for a model-friendly context block.

   Open source →

Your coding agent can read these notes before it upgrades. Set up the MCP server →

• akashipub.devProvider-neutral core for the Akashi agent framework: streaming tool-loop agents, typed tools, sealed message/event unions, multi-agent and durable-execution seams. Pure Dart.
• akashi_anthropicpub.devAnthropic (Claude) provider adapter for the Akashi agent framework: wraps anthropic_sdk_dart behind Akashi's LanguageModel contract, normalizing text/thinking/tool_use blocks.
• akashi_gemmapub.devOn-device LanguageModel for the Akashi agent framework, backed by flutter_gemma: run agents fully offline, or as the primary in a cloud FallbackModel chain.
• akashi_googlepub.devGoogle Gemini provider adapter for the Akashi agent framework: wraps googleai_dart behind Akashi's LanguageModel contract, with structured output, tool calls, and embeddings.