A multi-mode AI conversational system focused on humanities and social sciences, powered by Gemini 2.5 and Qwen.
Live: c4rows.com
This is a personal practice project exploring multi-agent architectures, RAG pipelines, and tool-calling patterns with Spring AI.
CrossRow provides four distinct chat modes, each tailored to different user needs:
| Mode | Description |
|---|---|
| Preferred | Direct multimodal chat with user's preferred model (Gemini or Qwen) |
| Auto | AI evaluates query complexity and routes to the appropriate model — Qwen for simple tasks, Gemini for complex ones |
| Agent | ReAct agent with tool calling — can search the web, generate images, ask clarifying questions, and retrieve domain knowledge |
| Expert | Multi-agent mode — an orchestrator routes queries to a specialized expert (philosophy, psychology, or sociology), each with domain-specific prompts and knowledge |
28 curated Markdown documents across three domains:
- Philosophy (12): Stoicism, Existentialism, Buddhism, Taoism, Confucianism, Kantianism, etc.
- Psychology (8): CBT, Psychoanalysis, Jungian Psychology, Adlerian Psychology, etc.
- Sociology (8): Marxism, Bourdieu, Foucault, Frankfurt School, Feminism, etc.
Documents are indexed into Elasticsearch at startup using hybrid retrieval — BM25 keyword search (with IK Chinese tokenizer) combined with KNN vector search (Vertex AI embeddings, 768-dim). Results are deduplicated and filtered by a dynamic score threshold.
| Tool | Description |
|---|---|
searchWeb |
Brave Search API, returns top 5 results |
generateImage |
gemini-2.5-flash(nano banana), supports multiple styles |
askHuman |
Pauses execution to ask the user for clarification |
terminate |
Ends the agent loop with a final answer |
updateUserMemory |
Persists user preferences/facts to Elasticsearch for long-term recall |
retrievePhilosophy/Psychology/Sociology |
Domain-specific hybrid RAG retrieval |
getCurrentTime |
Returns current timestamp |
calculator |
Basic arithmetic operations |
- Short-term: Redis-backed
ChatMemoryper session, with async compression — when token count exceeds thresholds, older messages are summarized by Qwen and merged into a compact system message. - Long-term: User-specific facts stored in Elasticsearch via
updateUserMemorytool, accessible across sessions.
Spring AI advisor chain applied to every LLM call:
| Advisor | Purpose |
|---|---|
PromptInjectionGuard |
Regex + Base64 + typoglycemia detection for injection attempts |
SimpleAuth |
Validates user exists in database |
SimpleQuota |
Enforces daily chat/agent usage limits per role |
ChatMemory |
Injects conversation history into prompts |
MyLog |
Logs request/response and token usage |
The project includes a benchmark suite (evaluation/) using three recognized methodologies:
| Benchmark | Framework | What it tests |
|---|---|---|
| RAGAS | RAGAS — 500+ citations | RAG retrieval quality: faithfulness (are answers grounded in retrieved docs?) and answer relevancy (are answers on-topic?) |
| LLM-as-Judge: Routing | Zheng et al. 2023 — 3000+ citations | Expert routing accuracy across 35 test cases including boundary questions |
| LLM-as-Judge: Quality | Same methodology | Response quality on 5 dimensions: relevance, professionalism, completeness, clarity, empathy |
| Dimension | Score | Grade |
|---|---|---|
| Routing Accuracy | 97.1% | A |
| Response Quality | 93.9 / 100 | A |
| RAG Quality | 79.3% | C |
- Routing: 35 test cases (10 philosophy, 10 psychology, 10 sociology, 5 boundary). Only 1 misrouted.
- Response Quality: Gemini judges each response on a 1-5 scale across 5 dimensions. Average 4.7/5.
- RAG Quality: Faithfulness 0.84 (low hallucination), answer relevancy 0.73 (philosophy domain retrieval is the weakest at 0.61 — documents are organized by school rather than by topic, causing mismatch on cross-school questions).
| Layer | Technology |
|---|---|
| Backend | Java 21, Spring Boot 3.2, Spring AI 1.1 |
| Frontend | React 19, Vite 7, Tailwind CSS 4 |
| LLM | Vertex AI Gemini 2.5, Qwen (via DashScope) |
| Embeddings | Vertex AI text-embedding (768-dim) |
| Database | PostgreSQL 17 (pgvector), Redis 7 |
| Search | Elasticsearch 8.17 (IK analyzer, hybrid BM25 + KNN) |
| Storage | Google Cloud Storage (images, user backgrounds) |
| Auth | JWT |
| Deploy | Docker Compose, GitHub Actions, Nginx |
| Evaluation | RAGAS, LLM-as-Judge (Gemini), Python |
