ask-llm-providers 0.1.5 → 0.1.6

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 98e73dfa923583db079237386caab024766c6613b202623b73952d3233ddf3ae
-  data.tar.gz: 0aeef8d5eface2a50ffa4365c1d8e53fb47de3d0be54eb88c4ba53dea6fecfff
+  metadata.gz: cf20ba02f9ecd52cdbc18f93ff8a22f75e9d33d845dd80ab7880d35454cd27c0
+  data.tar.gz: 982162bc42de9c71dfe83dabdd40b69a95b618ead9a056ffcafcaee8c77a5bd7
 SHA512:
-  metadata.gz: 94dd687fc071961fe2495e61788861d676470b42e7352ff5f2d9c53455fe090d9d93c1b122a7b1d769f347d81b7059b44fffb18235fa7c66910b36b193b9f890
-  data.tar.gz: 88d1dbcb9f4fb92d8a447352c725c6e27e89bbc29f46f7d83e96856c9288ec18bcbda236b04e6e7eb85df901187e3f5fcef56c464f5b101b3a12a53534dc2f20
+  metadata.gz: f4a361c8bf519e443793c31c4f21dc4f1a5872b4b64f1f7483499f22328a4db091fa6b4e287c0e56ece26777a5ea00e6a3a099c428def8a1cd91a5f61d88a74b
+  data.tar.gz: 93d0092ed08a53a12be9e822e7f8b4dc889695c39cd3f82f6167422110c8318c0e923b7f1cd1b00aa7104d5fb9f947490a43217ae35af703c570e7a5d8eb918a

data/lib/ask/llm/version.rb

@@ -2,6 +2,6 @@
 
 module Ask
   module LLM
-    VERSION = "0.1.5"
+    VERSION = "0.1.6"
   end
 end

data/lib/ask/skills/providers.model_select/SKILL.md

@@ -0,0 +1,164 @@
+---
+name: providers.model_select
+description: How to select the right LLM model for a task — balancing cost, capability, latency, and context window
+---
+
+Use this skill when choosing an LLM model for a specific task. The ask-rb
+ecosystem uses `Ask::ModelCatalog` to resolve models and check capabilities.
+
+## Step 1: Classify the Task
+
+Determine what kind of task you're solving:
+
+| Task Type | Examples | Key Requirements |
+|-----------|----------|-----------------|
+| **Simple chat** | Q&A, summarization, translation | Speed, low cost |
+| **Code generation** | Write functions, review PRs | Strong coding, large context |
+| **Reasoning/analysis** | Debugging, architecture, planning | Deep reasoning, structured output |
+| **Structured extraction** | Parse logs, extract data | JSON mode, function calling |
+| **Vision/multimodal** | Screenshot analysis, document OCR | Image input support |
+| **Long document** | Analyze 100+ page docs | Large context window (200K+) |
+| **Embeddings** | Semantic search, RAG | Embedding model, dimensions |
+
+## Step 2: Query the Model Catalog
+
+Access available models through the catalog:
+
+```ruby
+# List all models
+Ask::ModelCatalog.all
+
+# Filter by capability
+Ask::ModelCatalog.chat_models
+Ask::ModelCatalog.by_provider("openai")
+Ask::ModelCatalog.by_family("gpt")
+Ask::ModelCatalog.embedding_models
+```
+
+Find a specific model by ID:
+
+```ruby
+model = Ask::ModelCatalog.find("gpt-4o")
+model.context_window    # => 128000
+model.max_output_tokens # => 16384
+model.supports?(:function_calling) # => true
+model.capabilities      # => ["function_calling", "structured_output", "reasoning", "vision"]
+model.modalities        # => { input: ["text", "image"], output: ["text"] }
+```
+
+If the catalog doesn't have the model you need, refresh:
+
+```ruby
+Ask::ModelCatalog.refresh!
+```
+
+## Step 3: Evaluate Cost vs Capability
+
+Use pricing data from the catalog:
+
+```ruby
+model = Ask::ModelCatalog.find("gpt-4o")
+pricing = model.pricing.dig(:text_tokens, :standard)
+pricing[:input_per_million]   # $ per 1M input tokens
+pricing[:output_per_million]  # $ per 1M output tokens
+```
+
+**Cost comparison (approximate):**
+
+| Tier | Models | Cost/M tokens (in) | Best For |
+|------|--------|-------------------|----------|
+| **Frontier** | GPT-4o, Claude 4 Sonnet, Gemini 2.5 Pro | $3-15 | Complex reasoning, code generation |
+| **Fast/Cheap** | GPT-4o-mini, Claude 4 Haiku, Gemini 2.5 Flash | $0.15-1.00 | Simple chat, extraction, classification |
+| **Reasoning** | o3, o4-mini, DeepSeek R1 | $2-10 | Deep analysis, math, multi-step tasks |
+| **Specialized** | Embedding, image, audio models | Varies | Non-chat tasks |
+
+## Step 4: Match Capabilities to Task Requirements
+
+Check if a model supports the features you need:
+
+```ruby
+model.supports?(:function_calling) # For tool use
+model.supports?(:structured_output) # For JSON mode
+model.supports?(:vision)           # For image analysis
+model.supports?(:reasoning)        # For complex reasoning
+```
+
+**Capability requirements by task:**
+
+| Need | Check | Fallback |
+|------|-------|----------|
+| Tool calling | `supports?(:function_calling)` | Use text instruction instead |
+| JSON output | `supports?(:structured_output)` | Prompt-engineering |
+| Image processing | `modalities[:input].include?("image")` | Describe image in text |
+| Audio processing | `modalities[:input].include?("audio")` | Transcribe first |
+| Deep reasoning | `supports?(:reasoning)` | Chain-of-thought prompting |
+
+## Step 5: Consider Context Window Requirements
+
+Choose context window based on your input size:
+
+```ruby
+model.context_window  # total tokens the model can process
+```
+
+**Guidelines:**
+- **8K-16K** — Simple Q&A, short conversations
+- **32K-64K** — Code review, medium documents, multi-turn conversations
+- **100K-200K** — Large codebases, long documents, RAG with many chunks
+- **1M-2M** — Gemini 2.5 Pro, Gemini 2.0 Flash for massive documents
+
+Be aware that large context windows increase latency and cost even if you don't
+use them all.
+
+## Step 6: Pick the Right Embedding Model
+
+For RAG and semantic search:
+
+```ruby
+Ask::ModelCatalog.embedding_models
+```
+
+**Recommendations:**
+- **General purpose**: `text-embedding-3-large` (256-3072 dims)
+- **Best accuracy**: `text-embedding-3-large` with 3072 dimensions
+- **Fast/Cheap**: `text-embedding-3-small` (512 dimensions)
+- **Multilingual**: `text-embedding-3-small` (supports 100+ languages)
+
+## Decision Tree
+
+```
+Task Type?
+├── Simple chat / extraction
+│   └── Fast model (GPT-4o-mini, Claude 4 Haiku)
+│       → Cheapest adequate model
+├── Code generation / review
+│   └── Frontier model (GPT-4o, Claude 4 Sonnet)
+│       → Needs function calling + max capability
+├── Deep reasoning / debugging
+│   └── Reasoning model (o4-mini, DeepSeek R1, o3)
+│       → Needs chain-of-thought + analysis
+├── Long document analysis
+│   └── Large context (Gemini 2.5 Pro 1M, GPT-4o)
+│       → Needs context window > input size
+├── Multimodal (image/video)
+│   └── Vision-capable (GPT-4o, Claude 4 Sonnet, Gemini 2.5)
+│       → Check modalities[:input] includes image
+├── Embeddings / RAG
+│   └── text-embedding-3-large / small
+│       → Not a chat model
+└── Audio / Voice
+    └── GPT-4o-audio, Gemini Audio
+        → Check modalities[:output] includes audio
+```
+
+## Provider Selection
+
+Consider provider reliability and features:
+
+| Provider | Strengths | Weaknesses |
+|----------|-----------|------------|
+| **OpenAI** | Best tool calling, broad model range | Higher cost for frontier |
+| **Anthropic** | Excellent code, long context | Slower for simple tasks |
+| **Google Gemini** | Massive context (1M+), fast | Fewer integration tools |
+| **DeepSeek** | Cheap reasoning, open weights | Limited ecosystem |
+| **Ollama** | Local, free, private | Slow, no hosted offerings |

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: ask-llm-providers
 version: !ruby/object:Gem::Version
-  version: 0.1.5
+  version: 0.1.6
 platform: ruby
 authors:
 - Kaka Ruto
@@ -189,6 +189,7 @@ files:
 - lib/ask/provider/openai.rb
 - lib/ask/provider/opencode.rb
 - lib/ask/provider/opencode_go.rb
+- lib/ask/skills/providers.model_select/SKILL.md
 homepage: https://github.com/ask-rb/ask-llm-providers
 licenses:
 - MIT