Erupt AI RAG v2.1.0+
erupt-ai-rag is the knowledge-base extension of erupt-ai. Turn business documents into AI-searchable knowledge: upload documents → automatic chunking → vector embedding → semantic retrieval, all managed visually. The AI decides on its own when to query which knowledge base during conversations (Agentic RAG) — no manual mounting required.
Getting Started
- Add the dependency (requires erupt-ai):
<dependency>
<groupId>xyz.erupt</groupId>
<artifactId>erupt-ai-rag</artifactId>
<version>${erupt.version}</version>
</dependency>- After startup, the Knowledge Base menu group is added (the Embedding Model menu is provided by erupt-ai, under the AI Manager menu group).
Embedding Models
Go to AI → Embedding Model to configure the embedding model used for vectorization. 12 provider types are built in:
| Provider | Description |
|---|---|
| OpenAI Compatible | Any service compatible with the OpenAI Embedding API |
| Ollama | Locally deployed open-source embedding models |
| Gemini | Google Gemini Embedding |
| Qwen | Alibaba Cloud Qwen |
| GLM | Zhipu AI |
| SiliconFlow | SiliconFlow |
| Doubao | ByteDance Doubao (Volcano Engine) |
| Cohere | Cohere Embed |
| Voyage | Voyage AI |
| Jina | Jina Embeddings |
| Mistral | Mistral |
| OpenRouter | OpenRouter aggregation platform |
⚠️ The vector dimension cannot be changed after the first embedding; if a knowledge base's embedding model is changed, all documents must be re-embedded.
Vector Store
The vector store is deployment infrastructure and is configured via properties (leave blank to auto-select: the single persistent implementation on the classpath, otherwise in-memory storage):
erupt:
ai:
rag:
vector-store:
type: PGVECTOR # Options: QDRANT / MILVUS / PGVECTOR / REDIS / MEMORY
uri: postgresql://user:password@host:5432/db
api-key: # Optional authentication key| Type | uri format |
|---|---|
| QDRANT | host:6334 (gRPC, https:// prefix for TLS) |
| MILVUS | http://host:19530 |
| PGVECTOR | postgresql://user:password@host:5432/db (blank = reuse the application datasource) |
| REDIS | host:6379 (rediss:// prefix for TLS, blank = localhost) |
| MEMORY | No configuration needed; suitable for development |
Knowledge Bases & Documents
Go to AI → Knowledge Base to create a knowledge base, select an embedding model, and configure retrieval parameters:
| Parameter | Default | Description |
|---|---|---|
| Chunk Size | 500 | Max characters per chunk |
| Chunk Overlap | 50 | Characters shared by adjacent chunks |
| Top K | 5 | Number of chunks returned per retrieval |
| Min Score | 0.5 | Similarity threshold (0–1); lower-scored results are dropped |
💡 The Remark field is read by the AI to decide which knowledge base fits the current question — describe the content scope clearly.
Upload documents under a knowledge base (supports txt / md / markdown, or paste text directly). The system automatically chunks and embeds them, with status flowing: Pending → Embedding → Ready (errors are shown on failure and documents can be re-embedded).
The Retrieval Test row operation provides a visual interface — enter a question to verify recall results and similarity scores.
Agentic RAG
Knowledge base retrieval is exposed as AI Tools (listKnowledgeBases / searchKnowledgeBase). During conversations the AI decides autonomously: first list available knowledge bases, then run semantic retrieval against the appropriate one, preferring retrieved passages when answering. These tools are also governed by Role-Level Tool Authorization, giving fine-grained control over which roles may search which knowledge.
Contributors
YuePeng
Claude Fable 5.1