ragnar-cli 0.1.0.pre.4 → 0.1.0.pre.5

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: e0837b5d907d7b5336d938a4fa94c53b4dacdb96b56ffc753144dfaa4f476133
-  data.tar.gz: 7cd9d94241f8dc38a7dd4b3b2732966f21f9783b818087b30db3f03b5e6c2dfd
+  metadata.gz: 06c692710e5d5deb8cf5b8122050968deb459ad6feebd5bbdeabf63679a04d7c
+  data.tar.gz: dc8784c1b36aec7c473b9f93cc1929bdbd1d75ecec18d727c4e7892aeea6df30
 SHA512:
-  metadata.gz: 2a87f654f8502b292d3bfbea31c5f6bb5ba6f02638cd024e8efd623ec88c69f528c59ca4f2604437df9169ec4ad11ffcf4f6223441f9787b8d94ab312c149192
-  data.tar.gz: 133364ec6142c14ded8c58c7041aab342b290e9a196da54c9164f3dfc83d466410d173cb35aa5a5988b201dac251f34e441346f2debfcc0d8bb3e95e24476d1c
+  metadata.gz: cd6e88640e7629725fef32214f088d11362ddbcbbf52a90aa6d315ecc0b08c0d8f487501fd802221c9b647f1e22b55aa91a97f8eefa7046b38c2c49eff3d1bd3
+  data.tar.gz: b58174568e4b76d6cce4aeb19e7674cddb9f8c11c0c6e44a1da0e9f545263a3874b0f99cc16fdc3723e516ffb7c4ffee6660540c891cd595b805ded4f3ff0a49

data/README.md

@@ -2,6 +2,10 @@
 
 A complete Ruby implementation of Retrieval-Augmented Generation (RAG) pipeline using native Ruby ML/NLP gems.
 
+<p align="center">
+  <img src="/docs/assets/screenshot.png" alt="ragnar TUI" width="600">
+</p>
+
 ## Overview
 
 Ragnar provides a production-ready RAG pipeline for Ruby applications, integrating:
@@ -151,21 +155,41 @@ ragnar index ./documents \
   --chunk-overlap 100
 ```
 
-### 2. Train UMAP (Optional)
+### 2. Interactive Mode (TUI)
+
+Running `ragnar` with no arguments launches an interactive TUI powered by [ratatui](https://ratatui.rs/):
+
+```bash
+# Launch the TUI (default when no command given)
+ragnar
+
+# Or explicitly
+ragnar interactive
+```
+
+The TUI provides:
+- **Auto-completion** for commands and options
+- **Persistent history** across sessions
+- **Live output** — see indexing progress, query results, and topic analysis inline
+- **All CLI commands** available via `/command` syntax (e.g., `/index .`, `/umap train`, `/query "my question"`)
+- **`/verbose`** — toggle verbose mode to see query pipeline details (retrieval, reranking, context)
+- **`/profile`** — list or switch LLM profiles mid-session
+
+### 3. Train UMAP (Optional)
 
 Reduce embedding dimensions for faster search:
 
 ```bash
 # Train UMAP model (auto-adjusts parameters based on data)
-ragnar train-umap \
+ragnar umap train \
   --n-components 50 \
   --n-neighbors 15
 
 # Apply to all embeddings
-ragnar apply-umap
+ragnar umap apply
 ```
 
-### 3. Extract Topics
+### 4. Extract Topics
 
 Perform topic modeling to discover themes in your indexed documents:
 
@@ -194,7 +218,7 @@ The HTML export includes:
 - **Topic Bubbles**: Interactive bubble chart showing topic sizes and coherence
 - **Embedding Scatter Plot**: Visualization of all documents in embedding space, colored by cluster
 
-### 4. Query the System
+### 5. Query the System
 
 ```bash
 # Basic query
@@ -226,7 +250,7 @@ When using `--verbose` or `-v`, you'll see:
 6. **Response Generation**: The final LLM prompt and response
 7. **Final Results**: Confidence score and source attribution
 
-### 5. Check Statistics
+### 6. Check Statistics
 
 ```bash
 ragnar stats
@@ -290,81 +314,112 @@ Example `.ragnar.yml` file:
 ```yaml
 # Storage paths (all support ~ expansion)
 storage:
-  database_path: "~/.cache/ragnar/database"    # Vector database location
-  models_dir: "~/.cache/ragnar/models"         # Downloaded model files
-  history_file: "~/.cache/ragnar/history"      # Interactive mode history
+  database_path: "~/.cache/ragnar/database"
+  models_dir: "~/.cache/ragnar/models"
+  history_file: "~/.cache/ragnar/history"
 
 # Embedding configuration
 embeddings:
-  model: jinaai/jina-embeddings-v2-base-en    # Embedding model to use
-  chunk_size: 512                              # Tokens per chunk
-  chunk_overlap: 50                            # Token overlap between chunks
+  model: jinaai/jina-embeddings-v2-base-en
+  chunk_size: 512
+  chunk_overlap: 50
 
 # UMAP dimensionality reduction
 umap:
-  reduced_dimensions: 64                       # Target dimensions (2-100)
-  n_neighbors: 15                              # UMAP neighbors parameter
-  min_dist: 0.1                                # UMAP minimum distance
-  model_filename: umap_model.bin              # Saved model filename
+  reduced_dimensions: 64
+  n_neighbors: 15
+  min_dist: 0.1
+  model_filename: umap_model.bin
 
-# LLM configuration
+# LLM profiles — switch between local and cloud models
 llm:
-  default_model: TheBloke/TinyLlama-1.1B-Chat-v1.0-GGUF
-  default_gguf_file: tinyllama-1.1b-chat-v1.0.q4_k_m.gguf
+  default_profile: red_candle
+  profiles:
+    red_candle:
+      provider: red_candle
+      model: MaziyarPanahi/Qwen3-4B-GGUF
+    opus:
+      provider: anthropic
+      model: claude-opus-4-6
+      api_key: sk-ant-...           # or set ANTHROPIC_API_KEY env var
+    sonnet:
+      provider: anthropic
+      model: claude-sonnet-4-6
+    ollama:
+      provider: ollama
+      model: llama3.1:8b
 
 # Query processing
 query:
-  top_k: 3                      # Number of documents to retrieve
-  enable_query_rewriting: true  # Use LLM to improve queries
+  top_k: 3                          # Number of documents to retrieve
+  enable_query_rewriting: true       # Use LLM to improve queries
+  enable_reranking: true             # Cross-encoder reranking (disable for small corpora)
+  reranker_model: BAAI/bge-reranker-base  # Reranker model
 
 # Interactive mode
 interactive:
-  prompt: 'ragnar> '            # Command prompt
-  quiet_mode: true              # Suppress verbose output
+  prompt: 'ragnar> '
+  quiet_mode: true
 
 # Output settings
 output:
-  show_progress: true           # Show progress bars during indexing
+  show_progress: true
 ```
 
-### Viewing Configuration
+### LLM Profiles
+
+Profiles let you switch between LLM providers without editing config. Use local models for development and cloud models for production quality:
 
-Check current configuration:
 ```bash
-# Show all configuration settings
-ragnar config
+# Use a specific profile for a single command
+ragnar --profile opus query "What is our security policy?"
 
-# Show LLM model information
-ragnar model
+# In TUI mode, switch profiles mid-session
+ragnar> /profile           # List available profiles
+ragnar> /profile opus      # Switch to Opus
+ragnar> /profile red_candle  # Switch back to local
 ```
 
-In interactive mode:
+Ragnar supports any [RubyLLM](https://rubyllm.com/) provider: `red_candle` (local), `anthropic`, `openai`, `ollama`, and more. API keys can be set per-profile in the config or via environment variables (`ANTHROPIC_API_KEY`, `OPENAI_API_KEY`, etc.).
+
+### Viewing Configuration
+
 ```bash
-ragnar interactive
-ragnar> config    # Show configuration
-ragnar> model     # Show model details
+ragnar config     # Show all settings including active profile
+ragnar model      # Show LLM model information
+ragnar profile    # List all LLM profiles
+```
+
+In interactive mode (launch with `ragnar`):
+```
+ragnar> /config    # Show configuration
+ragnar> /profile   # List profiles
 ```
 
 ### Environment Variables
 
 Configuration values can be overridden with environment variables:
 - `XDG_CACHE_HOME` - Override default cache directory (~/.cache)
+- `ANTHROPIC_API_KEY` - Anthropic API key (used by anthropic profiles)
+- `OPENAI_API_KEY` - OpenAI API key (used by openai profiles)
 
 ### Supported Models
 
+**LLM Providers** (via RubyLLM):
+- `red_candle` — Local GGUF models (default): `MaziyarPanahi/Qwen3-4B-GGUF`, `MaziyarPanahi/Qwen3-8B-GGUF`
+- `anthropic` — Claude models: `claude-opus-4-6`, `claude-sonnet-4-6`
+- `openai` — GPT models: `gpt-4o`, `gpt-4o-mini`
+- `ollama` — Local Ollama models: `llama3.1:8b`, `mistral:7b`
+- Any other [RubyLLM provider](https://rubyllm.com/providers/)
+
 **Embedding Models** (via red-candle):
 - `jinaai/jina-embeddings-v2-base-en` (default, 768 dimensions)
 - `BAAI/bge-base-en-v1.5`
 - `sentence-transformers/all-MiniLM-L6-v2`
 
-**LLM Models** (via red-candle, GGUF format):
-- `TheBloke/TinyLlama-1.1B-Chat-v1.0-GGUF` (default, fast)
-- `TheBloke/Qwen2.5-1.5B-Instruct-GGUF`
-- `TheBloke/phi-2-GGUF`
-
-**Reranker Models** (via red-candle):
-- `BAAI/bge-reranker-base`
-- `cross-encoder/ms-marco-MiniLM-L-6-v2`
+**Reranker Models** (via red-candle, configurable):
+- `BAAI/bge-reranker-base` (default, XLM-RoBERTa)
+- `cross-encoder/ms-marco-MiniLM-L-12-v2` (smaller, BERT-based)
 
 ## Advanced Usage
 
@@ -586,5 +641,7 @@ This project integrates several excellent Ruby gems:
 - [red-candle](https://github.com/assaydepot/red-candle) - Ruby ML/LLM toolkit
 - [lancelot](https://github.com/scientist-labs/lancelot) - Lance database bindings
 - [clusterkit](https://github.com/scientist-labs/clusterkit) - UMAP and clustering implementation
+- [thor-interactive](https://github.com/scientist-labs/thor-interactive) - Interactive TUI for Thor CLIs
+- [ratatui_ruby](https://github.com/nicholasgasior/ratatui-ruby) - Ratatui terminal UI bindings
 - [parsekit](https://github.com/scientist-labs/parsekit) - Content extraction
 - [baran](https://github.com/moeki0/baran) - Text splitting utilities

data/lib/ragnar/cli.rb

@@ -1,4 +1,5 @@
 require_relative "cli_visualization"
+require_relative "cli_umap"
 require_relative "config"
 require "thor/interactive"
 require "stringio"
@@ -9,11 +10,16 @@ module Ragnar
     include CLIVisualization
     include Thor::Interactive::Command
 
+    default_command :interactive
+
+    class_option :profile, type: :string, aliases: "-p", desc: "LLM profile to use (e.g., red_candle, opus, sonnet)"
+
     # Configure interactive mode
     configure_interactive(
       prompt: Config.instance.interactive_prompt,
       allow_nested: false,
       history_file: Config.instance.history_file,
+      ui_mode: :tui,
       default_handler: proc do |input, thor_instance|
         puts "[DEBUG] Default handler called: #{input}" if ENV["DEBUG"]
 
@@ -37,6 +43,7 @@ module Ragnar
     class_variable_set(:@@cached_llm_manager, nil)
     class_variable_set(:@@cached_query_processor, nil)
     class_variable_set(:@@cached_db_path, nil)
+    class_variable_set(:@@verbose_mode, false)
 
     desc "index PATH", "Index text files from PATH (file or directory)"
     option :db_path, type: :string, desc: "Path to Lance database (default from config)"
@@ -87,83 +94,8 @@ module Ragnar
       end
     end
 
-    desc "train-umap", "Train UMAP model on existing embeddings"
-    option :db_path, type: :string, desc: "Path to Lance database (default from config)"
-    option :n_components, type: :numeric, default: 50, desc: "Number of dimensions for reduction"
-    option :n_neighbors, type: :numeric, default: 15, desc: "Number of neighbors for UMAP"
-    option :min_dist, type: :numeric, default: 0.1, desc: "Minimum distance for UMAP"
-    option :model_path, type: :string, desc: "Path to save UMAP model"
-    def train_umap
-      say "Training UMAP model on embeddings...", :green
-
-      config = Config.instance
-      # Use model_path from options if provided, otherwise use config models_dir
-      model_path = if options[:model_path]
-        options[:model_path]
-      else
-        File.join(config.models_dir, "umap_model.bin")
-      end
-      
-      processor = UmapProcessor.new(
-        db_path: options[:db_path] || config.database_path,
-        model_path: model_path
-      )
-
-      begin
-        stats = processor.train(
-          n_components: options[:n_components] || 50,
-          n_neighbors: options[:n_neighbors] || 15,
-          min_dist: options[:min_dist] || 0.1
-        )
-
-        say "\nUMAP training complete!", :green
-        say "Embeddings processed: #{stats[:embeddings_count]}"
-        say "Original dimensions: #{stats[:original_dims]}"
-        say "Reduced dimensions: #{stats[:reduced_dims]}"
-        say "Model saved to: #{processor.model_path}"
-      rescue => e
-        say "Error during UMAP training: #{e.message}", :red
-        exit 1
-      end
-    end
-
-    desc "apply-umap", "Apply trained UMAP model to reduce embedding dimensions"
-    option :db_path, type: :string, desc: "Path to Lance database (default from config)"
-    option :model_path, type: :string, desc: "Path to UMAP model"
-    option :batch_size, type: :numeric, default: 100, desc: "Batch size for processing"
-    def apply_umap
-      config = Config.instance
-      model_path = if options[:model_path]
-        options[:model_path]
-      else
-        File.join(config.models_dir, "umap_model.bin")
-      end
-
-      unless File.exist?(model_path)
-        say "Error: UMAP model not found at: #{model_path}", :red
-        say "Please run 'train-umap' first to create a model.", :yellow
-        exit 1
-      end
-
-      say "Applying UMAP model to embeddings...", :green
-
-      processor = UmapProcessor.new(
-        db_path: options[:db_path] || config.database_path,
-        model_path: model_path
-      )
-
-      begin
-        stats = processor.apply(batch_size: options[:batch_size] || 100)
-
-        say "\nUMAP application complete!", :green
-        say "Embeddings processed: #{stats[:processed]}"
-        say "Already processed: #{stats[:skipped]}"
-        say "Errors: #{stats[:errors]}" if stats[:errors] > 0
-      rescue => e
-        say "Error applying UMAP: #{e.message}", :red
-        exit 1
-      end
-    end
+    desc "umap SUBCOMMAND ...ARGS", "UMAP dimensionality reduction commands"
+    subcommand "umap", Umap
 
     desc "topics", "Extract and display topics from indexed documents"
     option :db_path, type: :string, desc: "Path to Lance database (default from config)"
@@ -172,9 +104,10 @@ module Ragnar
     option :export, type: :string, desc: "Export topics to file (json or html)"
     option :verbose, type: :boolean, default: false, aliases: "-v", desc: "Show detailed processing"
     option :summarize, type: :boolean, default: false, aliases: "-s", desc: "Generate human-readable topic summaries using LLM"
-    option :llm_model, type: :string, default: "TheBloke/TinyLlama-1.1B-Chat-v1.0-GGUF", desc: "LLM model for summarization"
-    option :gguf_file, type: :string, default: "tinyllama-1.1b-chat-v1.0.q4_k_m.gguf", desc: "GGUF file name for LLM model"
+    option :llm_model, type: :string, default: "MaziyarPanahi/Qwen3-4B-GGUF", desc: "LLM model for summarization"
+    option :gguf_file, type: :string, default: "Qwen3-4B.Q4_K_M.gguf", desc: "GGUF file name for LLM model"
     def topics
+      apply_profile!
       require_relative 'topic_modeling'
 
       say "Extracting topics from indexed documents...", :green
@@ -241,16 +174,12 @@ module Ragnar
         if options[:summarize] && topics.any?
           say "Generating topic summaries with LLM...", :yellow
           begin
-            require 'red-candle'
-
-            # Initialize LLM for summarization once
-            say "Loading model: #{options[:llm_model]}", :cyan if options[:verbose]
-            llm = Candle::LLM.from_pretrained(options[:llm_model], gguf_file: options[:gguf_file])
+            chat = LLMManager.instance.default_chat
 
             # Add summaries to topics
             topics.each_with_index do |topic, i|
               say "  Summarizing topic #{i+1}/#{topics.length}...", :yellow if options[:verbose]
-              topic.instance_variable_set(:@summary, summarize_topic(topic, llm))
+              topic.instance_variable_set(:@summary, summarize_topic(topic, chat))
             end
 
             say "Topic summaries generated!", :green
@@ -316,8 +245,10 @@ module Ragnar
     option :db_path, type: :string, desc: "Path to Lance database (default from config)"
     option :top_k, type: :numeric, default: 3, desc: "Number of top documents to use"
     option :verbose, type: :boolean, default: false, aliases: "-v", desc: "Show detailed processing steps"
+    option :rerank, type: :boolean, default: nil, desc: "Enable cross-encoder reranking (default from config)"
     option :json, type: :boolean, default: false, desc: "Output as JSON"
     def query(question)
+      apply_profile!
       puts "Debug - Query called with: #{question.inspect}" if ENV['DEBUG']
       puts "Debug - Options: #{options.inspect}" if ENV['DEBUG']
 
@@ -327,10 +258,11 @@ module Ragnar
       begin
         config = Config.instance
         result = processor.query(
-          question, 
-          top_k: options[:top_k] || config.query_top_k, 
-          verbose: options[:verbose] || false,
-          enable_rewriting: config.enable_query_rewriting?
+          question,
+          top_k: options[:top_k] || config.query_top_k,
+          verbose: options[:verbose] || @@verbose_mode,
+          enable_rewriting: config.enable_query_rewriting?,
+          enable_reranking: options[:rerank].nil? ? config.enable_reranking? : options[:rerank]
         )
         puts "Debug - Result keys: #{result.keys}" if ENV['DEBUG']
 
@@ -443,8 +375,12 @@ module Ragnar
       say "  Chunk overlap: #{config.chunk_overlap}"
       
       say "\nLLM:", :cyan
+      say "  Active profile: #{config.llm_profile_name}", :green
+      say "  Provider: #{config.llm_provider}"
       say "  Model: #{config.llm_model}"
-      say "  GGUF file: #{config.llm_gguf_file}"
+      if config.available_profiles.size > 1
+        say "  Available profiles: #{config.available_profiles.join(', ')}"
+      end
       
       say "\nUMAP:", :cyan
       say "  Reduced dimensions: #{config.get('umap.reduced_dimensions', Ragnar::DEFAULT_REDUCED_DIMENSIONS)}"
@@ -454,30 +390,77 @@ module Ragnar
       say "\nQuery:", :cyan
       say "  Top K: #{config.query_top_k}"
       say "  Query rewriting: #{config.enable_query_rewriting?}"
+      say "  Reranking: #{config.enable_reranking?}"
+      say "  Reranker model: #{config.reranker_model}" if config.enable_reranking?
     end
     
     desc "model", "Show current LLM model information"
     def model
       config = Config.instance
-      
+
       say "\nLLM Model Configuration:", :cyan
       say "-" * 40
-      
-      say "\nModel:", :green
-      say "  Repository: #{config.llm_model}"
-      say "  GGUF file: #{config.llm_gguf_file}"
-      
-      # Check if model files exist
-      model_path = File.join(config.models_dir, config.llm_gguf_file)
-      if File.exist?(model_path)
-        size_mb = (File.size(model_path) / 1024.0 / 1024.0).round(2)
-        say "\nModel file exists: #{model_path} (#{size_mb} MB)", :green
+
+      say "\nProfile: #{config.llm_profile_name}", :green
+      say "  Provider: #{config.llm_provider}"
+      say "  Model: #{config.llm_model}"
+
+      # Only show GGUF/local file info for local providers
+      if config.llm_provider == 'red_candle'
+        say "\nEmbedding Model: #{config.embedding_model}"
+
+        # Check if model files exist in HuggingFace cache
+        hf_cache = File.expand_path("~/.cache/huggingface/hub")
+        model_dir = config.llm_model.gsub("/", "--")
+        model_cache = File.join(hf_cache, "models--#{model_dir}")
+        if Dir.exist?(model_cache)
+          say "\nModel cached: #{model_cache}", :green
+        else
+          say "\nModel not yet downloaded (will download on first use)", :yellow
+        end
       else
-        say "\nModel file not found: #{model_path}", :yellow
-        say "Run 'ragnar query' to download automatically", :yellow
+        api_key = config.llm_api_key
+        env_key = case config.llm_provider
+                  when 'anthropic' then ENV['ANTHROPIC_API_KEY']
+                  when 'openai' then ENV['OPENAI_API_KEY']
+                  end
+        has_key = api_key || env_key
+        say "\nAPI key: #{has_key ? 'configured' : 'not set'}", has_key ? :green : :red
       end
     end
 
+    desc "profile [NAME]", "Show or switch LLM profile"
+    def profile(name = nil)
+      config = Config.instance
+
+      if name
+        begin
+          config.set_active_profile(name)
+          LLMManager.instance.clear_cache
+          say "Switched to profile: #{name}", :green
+          say "  Provider: #{config.llm_provider}"
+          say "  Model: #{config.llm_model}"
+        rescue ArgumentError => e
+          say e.message, :red
+        end
+      else
+        say "\nLLM Profiles:", :cyan
+        say "-" * 40
+        config.llm_profiles.each do |pname, pconfig|
+          active = pname == config.llm_profile_name ? " (active)" : ""
+          say "  #{pname}#{active}", active.empty? ? :white : :green
+          say "    Provider: #{pconfig['provider']}"
+          say "    Model: #{pconfig['model']}"
+        end
+      end
+    end
+
+    desc "verbose", "Toggle verbose mode on/off"
+    def verbose
+      @@verbose_mode = !@@verbose_mode
+      say "Verbose mode: #{@@verbose_mode ? 'on' : 'off'}", @@verbose_mode ? :green : :yellow
+    end
+
     desc "clear-cache", "Clear cached instances (useful in interactive mode)"
     def clear_cache_command
       clear_cache
@@ -665,6 +648,12 @@ module Ragnar
 
     private
 
+    def apply_profile!
+      return unless options[:profile]
+      Config.instance.set_active_profile(options[:profile])
+      LLMManager.instance.clear_cache
+    end
+
     # Cached instance helpers for interactive mode
     def get_cached_database(db_path = nil)
       # Use config default if no path provided
@@ -711,7 +700,7 @@ module Ragnar
     end
 
 
-    def summarize_topic(topic, llm)
+    def summarize_topic(topic, chat)
       # Get representative documents for context
       sample_docs = topic.representative_docs(k: 3)
 
@@ -728,7 +717,7 @@ module Ragnar
       PROMPT
 
       begin
-        summary = llm.generate(prompt).strip
+        summary = chat.ask(prompt).content.strip
         # Clean up common artifacts
         summary = summary.lines.first&.strip || "Related documents"
         summary = summary.gsub(/^(Summary:|Topic:|Documents:)/i, '').strip

data/lib/ragnar/cli_umap.rb

@@ -0,0 +1,86 @@
+# frozen_string_literal: true
+
+require "thor"
+
+module Ragnar
+  class CLI < Thor
+    class Umap < Thor
+      desc "train", "Train UMAP model on existing embeddings"
+      option :db_path, type: :string, desc: "Path to Lance database (default from config)"
+      option :n_components, type: :numeric, default: 50, desc: "Number of dimensions for reduction"
+      option :n_neighbors, type: :numeric, default: 15, desc: "Number of neighbors for UMAP"
+      option :min_dist, type: :numeric, default: 0.1, desc: "Minimum distance for UMAP"
+      option :model_path, type: :string, desc: "Path to save UMAP model"
+      def train
+        say "Training UMAP model on embeddings...", :green
+
+        config = Config.instance
+        model_path = if options[:model_path]
+          options[:model_path]
+        else
+          File.join(config.models_dir, "umap_model.bin")
+        end
+
+        processor = UmapProcessor.new(
+          db_path: options[:db_path] || config.database_path,
+          model_path: model_path
+        )
+
+        begin
+          stats = processor.train(
+            n_components: options[:n_components] || 50,
+            n_neighbors: options[:n_neighbors] || 15,
+            min_dist: options[:min_dist] || 0.1
+          )
+
+          say "\nUMAP training complete!", :green
+          say "Embeddings processed: #{stats[:embeddings_count]}"
+          say "Original dimensions: #{stats[:original_dims]}"
+          say "Reduced dimensions: #{stats[:reduced_dims]}"
+          say "Model saved to: #{processor.model_path}"
+        rescue => e
+          say "Error during UMAP training: #{e.message}", :red
+          exit 1
+        end
+      end
+
+      desc "apply", "Apply trained UMAP model to reduce embedding dimensions"
+      option :db_path, type: :string, desc: "Path to Lance database (default from config)"
+      option :model_path, type: :string, desc: "Path to UMAP model"
+      option :batch_size, type: :numeric, default: 100, desc: "Batch size for processing"
+      def apply
+        config = Config.instance
+        model_path = if options[:model_path]
+          options[:model_path]
+        else
+          File.join(config.models_dir, "umap_model.bin")
+        end
+
+        unless File.exist?(model_path)
+          say "Error: UMAP model not found at: #{model_path}", :red
+          say "Please run 'ragnar umap train' first to create a model.", :yellow
+          exit 1
+        end
+
+        say "Applying UMAP model to embeddings...", :green
+
+        processor = UmapProcessor.new(
+          db_path: options[:db_path] || config.database_path,
+          model_path: model_path
+        )
+
+        begin
+          stats = processor.apply(batch_size: options[:batch_size] || 100)
+
+          say "\nUMAP application complete!", :green
+          say "Embeddings processed: #{stats[:processed]}"
+          say "Already processed: #{stats[:skipped]}"
+          say "Errors: #{stats[:errors]}" if stats[:errors] > 0
+        rescue => e
+          say "Error applying UMAP: #{e.message}", :red
+          exit 1
+        end
+      end
+    end
+  end
+end