rails-ai-gateway 0.2.0 → 0.3.0

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: af35d6d1d275d95fdc5823eded924c98cd4e39e18e1f67388f1fd0f7e9f056f3
-  data.tar.gz: bc9d7f02956947396a2801b27199328aba3b38b1d9917e2e00d9a70a1fc46b11
+  metadata.gz: b09f7d65429d12dbd10f742e683c8a366302b465da1869f62f9c0ffb17be37a5
+  data.tar.gz: fb68e097b3437b2008f4352bc1c6905a4e57afd03be22cc6bb9ad709ea24e05d
 SHA512:
-  metadata.gz: 352f0608dbf281db2d14dcb452913cae5c758b4347d41a87bc38dc1fbc564792890a6b7f09edcab39741db71accdebb8ddeffb881a67c0b9900c358412da574f
-  data.tar.gz: 28780f6a22d71ac9c77ddc1a1f236fbbcf2e9fd12f8ed5412c495b7977ead1096c94a797e42642a1f1c8f87459077d5d8d8d35559645bb3f8d5652c7e85bcdd1
+  metadata.gz: 2f7c4bce60576871c5853bc6b7becd05752ac10cfad18b0a1d697587797381d94e800bacb87e9ba4f9983c35b2cb22c4ba28155828f629b562c5ff2bfe60e5a1
+  data.tar.gz: 40c43dae901d3a2f4b431a85c2bda10d51d40c3b3f602c95ec5d8f74ef6603676c9f41a905ec695e8a1975eb757dfeb5828f205f531ca611b00ebf9663251b9f

data/.yardopts

@@ -0,0 +1,6 @@
+--readme README.md
+--markup markdown
+--protected
+lib/**/*.rb
+-
+README.md

data/CHANGELOG.md

@@ -1,5 +1,20 @@
 # Changelog
 
+## 0.3.0 - 2026-09-21
+
+- Add durable input/output token totals grouped by public model.
+- Add per-route system prompt injection without prompt logging.
+- Add `RailsAIGateway::ENDPOINT`, `AI` inflection, and `RailsAIGateway` namespace alias.
+- Add one-key templates for popular OpenAI-compatible providers.
+- Add model capability metadata to routes, admin UI, and `/v1/models`.
+- Add provider logos, modal connection flow, and deterministic keyword query routing.
+- Add `models`, `model`, and `route_for` helpers for host Rails code.
+- Use same-origin referrer policy so browser form submissions pass Rails CSRF origin checks.
+- Add safe delete actions for providers and model routes in admin console.
+- Use Nebius Token Factory API endpoint in provider template.
+- Validate and forward OpenAI-compatible text, image, and audio message parts.
+- Filter query routing by declared model capabilities.
+
 ## 0.2.0 - 2026-09-20
 
 - Redesign admin UI as a responsive AI gateway operations console.

data/README.md

@@ -2,6 +2,7 @@
 
 [![Gem Version](https://badge.fury.io/rb/rails-ai-gateway.svg)](https://rubygems.org/gems/rails-ai-gateway)
 [![License: MIT](https://img.shields.io/badge/License-MIT-blue.svg)](LICENSE)
+[![RubyDoc](https://img.shields.io/badge/docs-RubyDoc-blue.svg)](https://www.rubydoc.info/gems/rails-ai-gateway)
 
 Put one OpenAI-compatible endpoint in front of your AI providers without running another
 service. Rails AI Gateway mounts inside your Rails app, stores configuration and request
@@ -15,6 +16,11 @@ metadata with ActiveRecord, and includes a Web UI for providers, model routes, a
 - Encrypted provider API keys using ActiveRecord encryption
 - Scoped gateway keys stored as SHA-256 digests
 - Request status, latency, attempt, and reported token-usage logs
+- Durable input/output token totals grouped by public model
+- Optional system prompt injection configured per model route
+- One-key provider templates for OpenAI, OpenRouter, Groq, DeepSeek, Mistral, Cerebras,
+  Nebius AI, and Perplexity
+- Explicit model capabilities exposed through UI and `/v1/models`
 - Configurable timeouts, body limits, fallback attempts, and network policy
 - SSRF protection with DNS validation and address pinning
 - Server-rendered, responsive, accessible admin UI
@@ -54,6 +60,9 @@ Web UI guides initial setup:
 2. Add public model route, such as `fast-chat` mapped to `gpt-5.6-luna`.
 3. Create gateway key and save token when shown. Raw token cannot be displayed again.
 
+Popular provider templates need only API key. Custom OpenAI-compatible provider form remains
+available for self-hosted and less common endpoints.
+
 Routes with same public model name form fallback chain. Lower priority runs first. Gateway
 retries connection failures before upstream accepts request and HTTP `429`, `500`, `502`,
 `503`, or `504`. Ambiguous timeouts and started streams are never retried.
@@ -63,7 +72,7 @@ retries connection failures before upstream accepts request and HTTP `429`, `500
 Runtime and security policy live in `config/initializers/rails_ai_gateway.rb`:
 
 ```ruby
-RailsAiGateway.configure do |config|
+RailsAIGateway.configure do |config|
   config.admin_controller = "ApplicationController"
   config.admin_authorization = ->(controller) {
     controller.current_user&.admin? == true
@@ -82,6 +91,16 @@ RailsAiGateway.configure do |config|
 end
 ```
 
+Rails acronym inflection exposes `RailsAIGateway` as the preferred constant while keeping
+`RailsAiGateway` compatible. Use the endpoint constant instead of repeating a URL path:
+
+```ruby
+OpenAI::Client.new(uri_base: "#{ENV.fetch("APP_URL")}#{RailsAIGateway::ENDPOINT}")
+```
+
+`RailsAIGateway::ENDPOINT` is `/ai/v1`, matching the generator's default mount. If host app
+changes mount path, build endpoint from that route instead.
+
 Admin requests are denied until `admin_authorization` returns exactly `true`. Provider
 definitions, model routes, gateway keys, and logs stay database-backed and editable through
 Web UI.
@@ -89,7 +108,7 @@ Web UI.
 Need another mount path? Change host route:
 
 ```ruby
-mount RailsAiGateway::Engine, at: "/gateway"
+mount RailsAIGateway::Engine, at: "/gateway"
 ```
 
 ## Make Requests
@@ -109,6 +128,32 @@ Available endpoints:
 
 OpenAI clients can use `http://localhost:3000/ai/v1` as base URL.
 
+Each model route can prepend an optional system message to chat requests. Injected prompts
+are stored on route and are never copied into request logs. Token dashboards use only usage
+reported by upstream providers; missing usage remains zero rather than being estimated.
+Routes declare capabilities from `text`, `vision`, `embedding`, `audio`, `video`, `tools`,
+and `reasoning`; clients receive merged capabilities in each `/v1/models` entry. Capability
+metadata does not add unsupported transport endpoints by itself.
+
+Chat requests support OpenAI-compatible multimodal content arrays containing `text`,
+`image_url` (HTTPS or image data URL), and `input_audio` (base64 WAV/MP3) parts. Gateway
+selects only routes advertising every required capability and forwards accepted parts without
+conversion. Request-size limit still applies to encoded media.
+
+Host Rails code can inspect active models without making an HTTP request:
+
+```ruby
+RailsAIGateway.models
+RailsAIGateway.models(capabilities: %w[vision tools])
+RailsAIGateway.models(provider: "OpenAI")
+RailsAIGateway.model("fast-chat")
+RailsAIGateway.route_for(model: "fast-chat", query: "Review this Ruby code", capabilities: %w[text tools])
+```
+
+Helpers return provider names and routing metadata, never provider API keys. `route_for`
+uses same literal keyword ranking as proxy: matching specialized routes first, then generic
+fallback routes.
+
 ## Development
 
 ```bash
@@ -117,6 +162,7 @@ cd rails-ai-gateway
 bundle install
 bundle exec ruby test/check.rb
 gem build rails_ai_gateway.gemspec
+rdoc --main README.md README.md lib
 ```
 
 SQLite integration runs by default. Run same suite against PostgreSQL with an empty test

data/app/assets/javascripts/rails_ai_gateway/admin.js

@@ -0,0 +1,25 @@
+document.addEventListener("DOMContentLoaded", () => {
+  const dialog = document.querySelector("#provider-connect-dialog");
+  if (!dialog || typeof dialog.showModal !== "function") return;
+
+  document.querySelectorAll("[data-provider-connect]").forEach((button) => {
+    button.addEventListener("click", () => {
+      const name = button.dataset.providerName;
+      dialog.querySelector("[data-dialog-template]").value = button.dataset.template;
+      dialog.querySelector("[data-dialog-logo]").src = button.dataset.logo;
+      dialog.querySelector("[data-dialog-key-label]").textContent = `${name} API Key`;
+      dialog.querySelector("[data-dialog-submit]").value = `Connect ${name}`;
+      dialog.querySelector("#connect-dialog-title").textContent = `Connect ${name}`;
+      dialog.showModal();
+      dialog.querySelector("[data-dialog-api-key]").focus();
+    });
+  });
+
+  dialog.querySelectorAll("[data-dialog-close]").forEach((button) => button.addEventListener("click", () => dialog.close()));
+  dialog.addEventListener("click", (event) => {
+    if (event.target === dialog) dialog.close();
+  });
+  dialog.addEventListener("close", () => {
+    dialog.querySelector("[data-dialog-api-key]").value = "";
+  });
+});