core/

AI Core Services (ai/core)

core 包提供了 DivineSense 全局使用的基础 AI 服务接口与实现。

Overview

These are foundational, well-tested services that form the cornerstone of upper-layer AI capabilities.

graph TD
    App[Application Layer / Agents] --> Core

    subgraph Core [AI Core Layer]
        LLM[llm: Unified Client]
        Embed[embedding: Vectorization]
        Rerank[reranker: Re-ranking]
        Ret[retrieval: Adaptive Retrieval]

        Ret --> LLM
        Ret --> Embed
        Ret --> Rerank
    end

    Core --> Providers[External Providers (OpenAI, DeepSeek, SiliconFlow, Z.AI)]

Subpackages

embedding (Vector Service)

Provides unified embedding generation interface for text vectorization.

• Unified Interface: Abstracts differences between providers (OpenAI, SiliconFlow, DashScope).
• Batch Processing: Supports concurrent batch operations with concurrency limiting (max 3 parallel).
• Chunking: Automatic document chunking with overlap for long content.
• Embedder: High-level API for memo embedding with automatic averaging pool.

llm (Large Language Model)

Unified Chat interface supporting streaming and function calling.

• Multi-Provider Support: DeepSeek, SiliconFlow, Z.AI (智谱), DashScope, OpenAI, OpenRouter, Ollama.
• Function Calling: Built-in tool/function calling support with low temperature for deterministic behavior.
• Streaming: Full streaming support with usage stats (TTFT, generation duration).
• Observability: Integrated metrics tracking (token usage, latency, cache hits).
• Warmup: Connection warmup for faster first request.

Key interfaces:

type Service interface {
    Chat(ctx context.Context, messages []Message) (string, *LLMCallStats, error)
    ChatStream(ctx context.Context, messages []Message) (<-chan string, <-chan *LLMCallStats, <-chan error)
    ChatWithTools(ctx context.Context, messages []Message, tools []ToolDescriptor) (*ChatResponse, *LLMCallStats, error)
    Warmup(ctx context.Context)
}

reranker (Re-ranking)

Used for the final step in RAG (Retrieval-Augmented Generation), re-ranking retrieved documents to improve accuracy.

• Provider: SiliconFlow (primary), generic OpenAI-compatible.
• Configurable: Can be disabled for fallback behavior.

retrieval (Adaptive Retrieval)

Implements intelligent retrieval with multiple strategies:

Retrieval Strategies

Advanced Features

• RRF Fusion: Reciprocal Rank Fusion for combining BM25 and vector results.
• Quality Assessment: Automatic result quality evaluation (High/Medium/Low).
• Dynamic Weighting: Adjusts BM25/Vector weights based on result quality.
• Fallback Logic: Automatic BM25 fallback when semantic quality is low.
• Smart Reranking: Conditional reranking based on query complexity.
• Time Optimization: Default 90-day window for vector search performance.