Advanced Information
🚀 Recent Advancements
Dense semantic vectors + sparse BM25 keyword vectors searched simultaneously in Qdrant. Catches exact-match terms that pure embedding search misses.
Cross-encoder reranking replaces LLM-based reranking. Attends jointly to query + document — +8–15% retrieval accuracy, highest nDCG@10 on BEIR.
API keys are Fernet-encrypted at rest. Keys are stripped from all server responses and never rendered into page source or DevTools.
Agentic reasoning-based retrieval — no embeddings or vector DB required. Uses tool-calling loops to locate and synthesise answers directly from document pages.
Multi-hop queries are decomposed into sub-questions. Hypothetical Document Embeddings (HyDE) expand sparse queries before retrieval.
Batch graph embedding and parallel doc-level indexing with skip-cache. Large documents index significantly faster with fewer redundant API calls.
Current Configuration
| Setting | Value |
|---|---|
| RAG Mode | Normal RAG |
| Preferred LLM | gemini |
| Embedding Model | Gemini Embedding API (768-dim) |
| Hybrid Search | Dense + BM25 Sparse |
| Reranker | Voyage AI Rerank 2.5 ✓ |
| Query Planning | HyDE + Decomposition ✓ |
| Vector Database | Qdrant Cloud |
| Document Collection | simple_rag_docs |
| Graph Collection | simple_rag_graph |
| Chunk Size | 1000 chars |
| Chunk Overlap | 200 chars |
| Results Count (Top K) | 5 |
| Embedding Cache | Enabled |
| Key Encryption | Fernet AES-128 at rest ✓ |
Retrieval Pipeline
HyDE expansion + sub-question decomposition
Dense (Gemini embeddings) + Sparse (BM25) searched in parallel
Voyage AI Rerank 2.5 cross-encoder re-scores the candidate pool
LLM generates grounded answer from top-ranked chunks
RAG Mode Comparison
📚 Normal RAG
- Speed: Fast
- Search: Hybrid dense + BM25
- Best for: Direct facts, document search
- Storage: Single vector collection
🕸️ Graph RAG
- Speed: Slower, more thorough
- Search: Hybrid + entity graph traversal
- Best for: Relationships, multi-hop reasoning
- Storage: Docs + graph collections
⚡ PageIndex
- Speed: Variable (agentic loop)
- Search: No vectors — tool-based page lookup
- Best for: Precise page-level extraction
- Storage: No vector DB required
🔒 Security
Encryption at Rest
API keys are encrypted with Fernet (AES-128-CBC + HMAC-SHA256) before being written to disk. The plaintext key never appears in the config file.
Zero Browser Exposure
Sensitive fields are stripped from every template context before rendering. Key values never appear in page source, network responses, or browser DevTools.
No Logging
Key values are never written to application logs — only field names are logged (e.g. "gemini_api_key updated"), never values.
Masked Form Fields
The setup form never pre-fills key inputs. Fields show "Already set — leave blank to keep" when a key exists, so values never transit to the browser.
About SimpleRAGx
SimpleRAGx is a multi-mode retrieval-augmented generation system with three complementary approaches to document Q&A, all backed by the same hybrid retrieval pipeline:
- Normal RAG: Fast hybrid semantic + keyword search for direct factual queries
- Graph RAG: Knowledge graph reasoning for complex relationships and multi-hop questions
- PageIndex: Agentic tool-based retrieval — no vector DB required, works directly on document pages
All modes share the same Qdrant backend, Gemini embeddings, and Voyage AI reranker. The retrieval pipeline (query planning → hybrid search → reranking → synthesis) is applied consistently across modes.