gleann Agent工作流

gleann

A lightweight, high-performance AI/RAG workspace and autonomous agent framework implemented in Go. Inspired by the Leann RAG backend architecture, designed for terminal environments.

🤖 Note: This project, including its documentation, was developed with the assistance of AI.

---

Key Features

• Context Field Theory (Φ Scoring): MCP search results are re-ranked using a multi-factor Φ score — recency (1hr half-life decay), frequency, graph proximity, and structural degree centrality.
• 10 File Read Modes: Smart file reading for LLM agents: map, signatures, entropy, diff, task, reference, aggressive, lines:N, auto, and full. Saves 60–90% of context-window tokens.
• Shell Output Compression: 95+ tool-specific regex patterns (Git, Go, NPM, Docker, Cargo) collapse noisy terminal output before it enters the LLM context window.
• 17 Agent & IDE Platforms: gleann install auto-detects and configures OpenCode, Claude Code, Cursor, Codex, Gemini CLI, Windsurf, Cline/Roo, Amp, Kiro, Amazon Q, Continue, Zed, Neovim, JetBrains, OpenClaw, Aider, and GitHub Copilot CLI.
• Token Gain Tracking: gleann_gain MCP tool reports cumulative token savings across a session, enabling budget-aware agent loops.
• Academic Vision, Full-Fledged Agent: Built on the shoulders of Leann's RAG architecture to create an autonomous assistant where LLM, vector/graph DBs, and plugins unite in one Go app.
• Zero-Config Extractive Summarization: High-density sentences are extracted algorithmically during build time, bypassing LLMs and enabling zero-latency "Smart Summaries".
• Flexible Intelligence (Local or Cloud): Run LLMs 100% locally via llama.cpp for total privacy, or connect to any OpenAI-compatible API for high-reasoning tasks.
• Advanced RAG (Faiss / HNSW & Graph DB): Indexes documents and code semantically (vector) and relationally (graph), not just via simple keyword matching.
• Smart Chunking (Tree-sitter): Intelligent AST-aware partitioning preserves the structural integrity of your code functions and classes.
• Graph-Augmented Search: Search results are enriched with callers/callees from the AST graph, giving LLMs structural code context alongside semantic matches.
• Impact Analysis: Blast radius analysis via BFS traversal — find all direct and transitive callers of any symbol and the files they belong to.
• Multi-Index Chat: Ask questions across multiple indexes simultaneously with gleann ask docs,code "question". Results are merged by relevance score.
• Conversations: Persistent conversation history with --continue, --continue-last, --title. Manage via gleann chat --list / --show / --delete.
• Roles & Format Control: Named system prompt roles (--role code, --role shell) and output format control (--format json, --format markdown). Custom roles in config.
• Markdown Rendering: Terminal markdown rendering via glamour. Disable with --raw.
• Word-wrap: Terminal-aware word wrapping with --word-wrap N for streaming output.
• LLM Title Summarization: Auto-generated conversation titles via LLM when no title is provided.
• Embedding Cache: Two-tier cache (L1: otter in-memory ≤50k vectors; L2: disk keyed by SHA-256). L2 hits are promoted to L1; unchanged chunks skip recompute entirely during rebuilds.
• Pipe-Friendly: Full stdin/pipe support (cat file | gleann ask index "review"), auto-raw mode when stdout is piped, --quiet for scripting.
• No-Cache / No-Limit: --no-cache skips conversation save, --no-limit removes token cap for unlimited output.
• .gleannignore: Gitignore-style patterns to exclude files during index builds.
• Config Management: gleann config show/path/edit/validate for easy configuration.
• Model Context Protocol (MCP) Server: A background service that bridges the gap between your local context and AI tools like Cursor or Claude Desktop.
• Long-term Memory (BBolt Blocks): Hierarchical short/medium/long-term memory that is automatically injected into every LLM query. Store facts with /remember, browse with /memories.
• OpenAI-Compatible Proxy: Drop-in replacement for OpenAI API — use any OpenAI SDK with model: "gleann/<index>" for instant RAG.
• Batch Query (MCP): gleann_batch_ask runs up to 10 questions concurrently against an index in a single round-trip.
• Rate Limiting & Timeouts: Per-IP token-bucket rate limiting (429) and per-endpoint context deadlines (504) protect the server in production.
• Retry Logic: Automatic exponential-backoff retry for transient LLM/embedding failures (503, 502, 429, connection refused).
• Background Maintenance: Scheduler auto-promotes memory blocks between tiers and prunes expired entries.
• A2A Protocol (Agent-to-Agent): Google's A2A protocol for agent discovery — other AI agents find and communicate with gleann via /.well-known/agent-card.json.
• Unified Memory API: Single POST /api/memory/ingest + POST /api/memory/recall interface that orchestrates block memory, knowledge graph, and vector search in parallel.
• Multimodal Detection: Automatically detects and uses multimodal Ollama models (Gemma4, Qwen3-VL, LLaVA) for processing images, audio, and video.
• Background Task Manager: Monitor long-running operations (indexing, memory consolidation) with progress tracking via GET /api/tasks.
• Auto-Bootstrap: gleann serve detects Ollama, selects models, and creates an initial config file without manual intervention.
• gleann setup --auto — Zero-Friction Onboarding Tool: Detects environment, pulls required models, and builds initial indexes via an automated workflow.
• Cross-Platform Service Management: gleann service install/start/stop/status manages a background server via systemd (Linux), launchd (macOS), or Task Scheduler (Windows).
• Auto Model Management: Missing models are automatically retrieved with progress tracking.
• Tiered Model Strategy: Defaults to lightweight models for fast initialization, with the ability to configure larger models for advanced use cases.
• Terminal User Interface (TUI): A keyboard-centric interface for interacting with indexed data and executing AI operations directly from the shell.