docs: add RAG advanced search guide and CLI commands

New docs/rag/advanced-search.md:
- Reranker (cross-encoder for +30-40% accuracy)
- Hybrid Search (BM25 + dense for +20% recall)
- HyDE (hypothetical document embeddings)
- Multi-Query (variant generation with RRF)
- Query Classifier (adaptive retrieval routing)
- Configuration reference for all features

Updated docs/reference/cli.md:
- Added RAG Operations section with rag-cli commands
- search, ingest, get-chunk, list, health
- Image search commands
- Updated AI-friendly summary

Navigation: Added RAG Advanced to LKAP menu

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: madeinoz67

docs/rag/advanced-search.md

@@ -0,0 +1,279 @@
+---
+title: "Advanced RAG Search"
+description: "Advanced retrieval techniques including reranking, hybrid search, HyDE, and multi-query"
+---
+
+# Advanced RAG Search
+
+The RAG system includes several advanced retrieval techniques that improve search accuracy and recall. These are enabled by default and can be configured via environment variables.
+
+## Overview
+
+| Feature | Purpose | Improvement |
+|---------|---------|-------------|
+| **Reranker** | Cross-encoder relevance scoring | +30-40% accuracy |
+| **Hybrid Search** | BM25 + dense vector fusion | +20% recall |
+| **HyDE** | Hypothetical document expansion | Better for short queries |
+| **Multi-Query** | Query variant generation | Covers more interpretations |
+| **Query Classifier** | Adaptive retrieval routing | Right strategy per query |
+
+## Search Pipeline
+
+```
+Query → Query Classifier → [HyDE?] → [Multi-Query?] → Hybrid Search → Reranker → Results
+            ↓                    ↓            ↓              ↓            ↓
+        Type: factual       Expand      3 variants     Dense+BM25   Cross-encoder
+            conceptual       short        merged         RRF         top-k
+```
+
+---
+
+## Reranker
+
+Cross-encoder reranking improves retrieval accuracy by 30-40% by using a more precise (but slower) model to score candidates.
+
+### How It Works
+
+1. **Initial retrieval**: Bi-encoder (fast, approximates relevance) returns top-20
+2. **Reranking**: Cross-encoder (slow, precise) scores each candidate
+3. **Final results**: Top-10 by cross-encoder score
+
+### When Reranking Helps
+
+- **High-precision requirements** - Need the most relevant results
+- **Complex queries** - Multiple concepts or ambiguous terms
+- **Production systems** - "Never skip reranking for production RAG"
+
+### Configuration
+
+```bash
+# Enable/disable reranking (default: true)
+MADEINOZ_KNOWLEDGE_RERANKER_ENABLED=true
+
+# Cross-encoder model
+MADEINOZ_KNOWLEDGE_RERANKER_MODEL=BAAI/bge-reranker-base
+
+# Candidates to rerank (default: 20)
+MADEINOZ_KNOWLEDGE_RERANKER_TOP_K=20
+
+# Final results after reranking (default: 10)
+MADEINOZ_KNOWLEDGE_RERANKER_FINAL_K=10
+
+# Provider: local, openrouter, cohere (default: local)
+MADEINOZ_KNOWLEDGE_RERANKER_PROVIDER=local
+```
+
+---
+
+## Hybrid Search
+
+Combines vector similarity (dense) with keyword matching (sparse/BM25) for better recall, especially for:
+
+- **Acronyms and proper nouns** - "GPT-4" vs "GPT" and "4"
+- **Exact phrase matching** - "machine learning" as phrase
+- **Rare terms** - Not well-represented in embedding space
+
+### How It Works
+
+1. **Dense retrieval**: Vector similarity via Qdrant
+2. **Sparse retrieval**: BM25 keyword matching via Qdrant text index
+3. **Fusion**: Reciprocal Rank Fusion (RRF) combines results
+
+### RRF Formula
+
+```
+score(d) = sum(1 / (k + rank)) for each result list
+k = 60 (dampens rank impact)
+```
+
+### Configuration
+
+```bash
+# Enable hybrid search (default: true)
+MADEINOZ_KNOWLEDGE_HYBRID_ENABLED=true
+
+# Weight for dense vs sparse (default: 0.7 = favor dense)
+MADEINOZ_KNOWLEDGE_HYBRID_ALPHA=0.7
+
+# RRF constant k (default: 60)
+MADEINOZ_KNOWLEDGE_HYBRID_RRF_K=60
+```
+
+---
+
+## HyDE (Hypothetical Document Embeddings)
+
+Generates a hypothetical answer to the query, then retrieves documents similar to that hypothetical document. Best for short, ambiguous queries.
+
+### How It Works
+
+1. **Query**: "login issues"
+2. **Hypothetical doc**: "Authentication errors may occur due to invalid credentials, expired sessions, or password reset requirements..."
+3. **Retrieve**: Find docs similar to hypothetical
+
+### When HyDE Helps
+
+- Short, ambiguous queries ("login issues")
+- Queries with little domain terminology
+- Documents are more verbose than queries
+
+### When to Skip HyDE
+
+- Long, specific queries (already contain good keywords)
+- Latency is critical (requires LLM call)
+- LLM might hallucinate domain terminology
+
+### Configuration
+
+```bash
+# Enable HyDE expansion (default: true)
+MADEINOZ_KNOWLEDGE_HYDE_ENABLED=true
+
+# Min query tokens to trigger HyDE (default: 10)
+MADEINOZ_KNOWLEDGE_HYDE_MIN_QUERY_TOKENS=10
+
+# Max tokens in hypothetical document (default: 200)
+MADEINOZ_KNOWLEDGE_HYDE_MAX_HYPOTHETICAL_TOKENS=200
+```
+
+### Cost
+
+- Adds LLM call per query (~$0.002 per query at 200 tokens)
+- For 10K queries/day: ~$600/month
+
+---
+
+## Multi-Query Variants
+
+Generates multiple query variants/rephrasings, retrieves for each, then merges results using RRF.
+
+### How It Works
+
+1. **Original query**: "How do I configure SPI?"
+2. **Variants**:
+   - "SPI configuration settings"
+   - "Set up Serial Peripheral Interface"
+   - "SPI master/slave setup guide"
+3. **Retrieve** for each variant
+4. **Merge** with RRF
+
+### When Multi-Query Helps
+
+- Complex queries with multiple interpretations
+- Queries that might match different terminology
+- Single query retrieval yields poor results
+
+### When to Skip
+
+- Simple, well-defined queries
+- Exact match queries (keywords, IDs)
+- Latency is critical
+
+### Configuration
+
+```bash
+# Enable multi-query (default: true)
+MADEINOZ_KNOWLEDGE_MULTI_QUERY_ENABLED=true
+
+# Min query length to trigger (default: 10)
+MADEINOZ_KNOWLEDGE_MULTI_QUERY_MIN_LENGTH=10
+
+# Number of variants to generate (default: 3)
+MADEINOZ_KNOWLEDGE_MULTI_QUERY_NUM_VARIANTS=3
+```
+
+---
+
+## Query Classifier
+
+Classifies query type and routes to appropriate retrieval strategy.
+
+### Query Types
+
+| Type | Description | Best Strategy |
+|------|-------------|---------------|
+| **factual** | Specific facts, exact matches | Keyword retriever |
+| **procedural** | How-to, step-by-step | Hierarchical retriever |
+| **conceptual** | Explanations, understanding | Vector retriever |
+| **comparative** | Comparing options | Multi-document |
+| **temporal** | Time-sensitive | Time-filtered |
+| **ambiguous** | Needs clarification | Ask clarification |
+
+### Classification Methods
+
+- **Rule-based** (fast, no LLM cost) - Default
+- **LLM-based** (more accurate, higher cost)
+- **Hybrid** (rules first, LLM for ambiguous)
+
+### Configuration
+
+```bash
+# Enable classification (default: true)
+MADEINOZ_KNOWLEDGE_QUERY_CLASSIFIER_ENABLED=true
+
+# Use LLM for classification (default: false = rule-based)
+MADEINOZ_KNOWLEDGE_QUERY_CLASSIFIER_USE_LLM=false
+```
+
+---
+
+## Complete Configuration Reference
+
+Add these to your `.env` file:
+
+```bash
+# Reranker
+MADEINOZ_KNOWLEDGE_RERANKER_ENABLED=true
+MADEINOZ_KNOWLEDGE_RERANKER_MODEL=BAAI/bge-reranker-base
+MADEINOZ_KNOWLEDGE_RERANKER_TOP_K=20
+MADEINOZ_KNOWLEDGE_RERANKER_FINAL_K=10
+MADEINOZ_KNOWLEDGE_RERANKER_PROVIDER=local
+
+# Hybrid Search
+MADEINOZ_KNOWLEDGE_HYBRID_ENABLED=true
+MADEINOZ_KNOWLEDGE_HYBRID_ALPHA=0.7
+MADEINOZ_KNOWLEDGE_HYBRID_RRF_K=60
+
+# HyDE
+MADEINOZ_KNOWLEDGE_HYDE_ENABLED=true
+MADEINOZ_KNOWLEDGE_HYDE_MIN_QUERY_TOKENS=10
+MADEINOZ_KNOWLEDGE_HYDE_MAX_HYPOTHETICAL_TOKENS=200
+
+# Multi-Query
+MADEINOZ_KNOWLEDGE_MULTI_QUERY_ENABLED=true
+MADEINOZ_KNOWLEDGE_MULTI_QUERY_MIN_LENGTH=10
+MADEINOZ_KNOWLEDGE_MULTI_QUERY_NUM_VARIANTS=3
+
+# Query Classifier
+MADEINOZ_KNOWLEDGE_QUERY_CLASSIFIER_ENABLED=true
+MADEINOZ_KNOWLEDGE_QUERY_CLASSIFIER_USE_LLM=false
+```
+
+---
+
+## Performance vs Accuracy Trade-offs
+
+| Configuration | Latency | Accuracy | Cost | Use Case |
+|--------------|---------|----------|------|----------|
+| All enabled | High | Best | High | Production, critical |
+| Reranker only | Medium | Good | Low | Balanced |
+| Hybrid only | Low | Good | Free | Speed priority |
+| All disabled | Fastest | Baseline | Free | Development |
+
+## Disabling Features
+
+For faster development iteration, disable advanced features:
+
+```bash
+# Fast development mode
+MADEINOZ_KNOWLEDGE_RERANKER_ENABLED=false
+MADEINOZ_KNOWLEDGE_HYBRID_ENABLED=false
+MADEINOZ_KNOWLEDGE_HYDE_ENABLED=false
+MADEINOZ_KNOWLEDGE_MULTI_QUERY_ENABLED=false
+```
+
+## Related Documentation
+
+- [RAG Quickstart](quickstart.md) - Basic search usage
+- [RAG Configuration](configuration.md) - Qdrant and Ollama setup
+- [RAG Troubleshooting](troubleshooting.md) - Common issues