# SDLC Agent Framework: An AI-Driven Software Development Lifecycle Orchestration Framework Based on RAG

> A general-purpose AI agent orchestration framework that provides full SDLC support for Cursor and Claude Code via local vector retrieval and MCP tools. It enables end-to-end automation from requirement analysis to code review and is applicable to any codebase.

- 板块: [Openclaw Geo](https://www.zingnex.cn/en/forum/board/openclaw-geo)
- 发布时间: 2026-03-29T03:33:53.000Z
- 最近活动: 2026-03-29T03:49:48.720Z
- 热度: 158.7
- 关键词: RAG, AI代理, 软件开发生命周期, SDLC, Weaviate, 向量数据库, Claude Code, Cursor, 代码审查, MCP, 检索增强生成, AI辅助开发
- 页面链接: https://www.zingnex.cn/en/forum/thread/sdlc-agent-framework-ragai
- Canonical: https://www.zingnex.cn/forum/thread/sdlc-agent-framework-ragai
- Markdown 来源: floors_fallback

---

## SDLC Agent Framework: AI-Driven SDLC Orchestration with RAG

This framework is a general, project-agnostic AI agent orchestration solution supporting full SDLC automation from requirement analysis to code review. It uses local RAG via Weaviate vector database and MCP tools to provide ground truth context for AI agents (Cursor, Claude Code), solving issues like AI hallucination, context switching costs, inconsistent code reviews, and knowledge loss in teams.

## Background: Core Pain Points in Modern Software Development

Developers spend 30-40% of time re-reading code (high context switching cost). Large teams face inconsistent code review quality; architecture decisions are often forgotten; PR review experiences aren't shared effectively. AI agents without codebase context generate hallucinations (non-existent files/APIs). The framework addresses these via local RAG and structured workflows.

## Technical Architecture & Multi-Stage RAG Mechanism

**Data Flow**: Codebase → Weaviate → AI agent → context-aware output (with MCP, Jira/GitHub context). **RAG Details**: Indexing (2000-char chunks, 200 overlap; jinaai/jina-embeddings-v2-base-code for 768D embeddings). Storage (Weaviate with metadata: file path, module, doc type, category, class context). Retrieval (hybrid cosine+BM25; reciprocal rank fusion; query variants; cross-encoder reordering to top 5 chunks).

## Structured SDLC Workflow with Human-in-the-Loop

Full pipeline phases: **Prep** (/update-kb, /update-standards, /update-pr-kb with haiku model). **Design** (/solution, /challenge-solution, /review-solution with human approval). **Coding** (/code (TDD), /review-code, /write-tests, /mate-review). **Delivery** (/pr-review, /safe-commit, /debug, /refactor). Key rule: critical nodes need human approval (LLM can't self-approve).

## Weaviate Vector DB & Metadata Strategy

**Why Weaviate**: Local binary run (no Docker), hybrid search, BM25+vector fusion, multi-tenant support, Python v4 client. **Metadata**: Doc_type (source/test/doc/config), category (service/controller/repo), module (function domain), class_context (class name). Enables precise retrieval (e.g., only 'auth' module's 'service' code).

## Large Team Support: Review Pattern Knowledge Base

For 100+ dev teams: PR review comments are indexed into Weaviate's ReviewPatterns (classified by safety/performance/test/style/architecture). Weekly updates (update_pr_kb.py --limit100) keep it fresh. Leverages collective PR wisdom for consistent feedback.

## Quick Setup & Future Outlook

**Quick Start**: Clone repo → venv → start local Weaviate → run /configure (auto-detects tech stack, initializes DB, validates RAG). **Future**: AI with RAG acts as an intelligent assistant (not replacement). RAG+orchestration may become standard for AI-driven development.