diff --git a/app/models/message.rb b/app/models/message.rb
index 2dc23b6..7898526 100644
--- a/app/models/message.rb
+++ b/app/models/message.rb
@@ -1,7 +1,7 @@
 class Message < ApplicationRecord
   acts_as_message tool_calls_foreign_key: :message_id
   has_many_attached :attachments
-  broadcasts_to ->(message) { "chat_#{message.chat_id}" }, inserts_by: :append, target: "messages"
+  broadcasts_to ->(message) { "chat_#{message.chat_id}" }, inserts_by: :append, target: "messages", partial: "messages/message"
 
   after_update_commit :broadcast_message_replacement, if: :assistant?
   before_save :calculate_cost, if: :should_calculate_cost?
diff --git a/docs/ideas/rag-chatbot.md b/docs/ideas/rag-chatbot.md
new file mode 100644
index 0000000..0e7b4c5
--- /dev/null
+++ b/docs/ideas/rag-chatbot.md
@@ -0,0 +1,182 @@
+# Idea: RAG Chatbot ("Ask about Ruby")
+
+A small chatbot on the homepage where visitors can learn about Ruby, augmented with our own content (Posts, Testimonials, Projects) via vector search.
+
+## Goals
+
+- Add a simple "Ask about Ruby" experience to the homepage
+- Gain experience with vector search using SQLite (no separate DB)
+- Reuse existing chat infrastructure (Chat/Message models, views, streaming)
+- Keep costs bounded and prevent abuse as a free general-purpose AI
+
+## UX
+
+- **Homepage**: a single input field ("Ask about Ruby...") with a send button — no full chat widget
+- **Homepage nav**: new "Chat" link alongside existing category buttons and "Community"
+- **`/chat` page**: reuses the existing beautiful chat view (from the remote template), requires login
+- **Single chat per user** — no chat list, no "new chat" button, no multi-tenant URLs exposed to users
+- **Auth gate**: hitting send on the homepage, or navigating to `/chat`, triggers GitHub OAuth if not signed in, then lands on `/chat` with the question submitted
+
+## Cost control — single rotating chat
+
+A single ever-growing chat would get expensive fast because RubyLLM's `acts_as_chat` sends full conversation history on every request.
+
+**Strategy:** one active chat per user, silently rotated:
+
+- Always show "the user's current chat" at `/chat`
+- If the current chat exceeds ~30 messages or is older than 24h, start a new one behind the scenes
+- Old chats stay in the DB for cost tracking but the user never sees them
+- From the user's perspective it feels like one continuous conversation
+
+## Topic restriction
+
+Restrict conversations to Ruby / our site content via the system prompt. No code-level topic detection — the LLM enforces this. If a question isn't about Ruby, the Ruby ecosystem, or content on our site, politely decline and redirect.
+
+## Architecture: RAG with SQLite vector search
+
+### Stack
+
+| Layer | Tool | Notes |
+|-------|------|-------|
+| Vector storage | `sqlite-vec` gem | SQLite extension, no separate DB |
+| Rails integration | `neighbor` gem | Andrew Kane — `has_neighbors`, `nearest_neighbors` |
+| Embeddings | `RubyLLM.embed` | Already have `ruby_llm` |
+| Chat + streaming | Existing `Chat`/`Message` + `ChatResponseJob` + Turbo Streams | Full reuse |
+
+Two new gems: `sqlite-vec`, `neighbor`. Everything else already in the project.
+
+### Flow
+
+```
+User question
+     │
+     ▼
+ ① RubyLLM.embed(question)  →  query vector (1536 floats)
+     │
+     ▼
+ ② ContentChunk.nearest_neighbors(:embedding, vector, distance: "cosine").first(8)
+     │
+     ▼
+ ③ Build system prompt with retrieved chunks + Ruby-only instructions
+     │
+     ▼
+ ④ chat.ask(question)  →  streamed response via existing Turbo Streams
+```
+
+### Data model
+
+**Two new tables:**
+
+```ruby
+# vec0 virtual table — stores only vectors + IDs
+create_virtual_table :content_embeddings, :vec0, [
+  "id text PRIMARY KEY NOT NULL",
+  "embedding float[1536] distance_metric=cosine"
+]
+
+# Metadata table — stores text chunks + polymorphic source reference
+create_table :content_chunks, id: { type: :string, default: -> { "uuid7()" } } do |t|
+  t.string :source_type, null: false  # "Post", "Testimonial", "Project"
+  t.string :source_id, null: false
+  t.text :content, null: false
+  t.string :title
+  t.timestamps
+end
+```
+
+### Content indexed
+
+| Model | Fields embedded |
+|-------|----------------|
+| **Post** (published) | `title + summary + content_or_url` |
+| **Testimonial** (published) | `heading + quote + body_text` |
+| **Project** (visible) | `name + description + topics` |
+
+### New concerns
+
+**`Embeddable`** — shared by Post, Testimonial, Project. Each model defines `embeddable_text`. Callbacks trigger `GenerateEmbeddingJob` on save.
+
+**`Chat::RagAugmentable`** — added to Chat. Provides `ask_with_rag(question, &block)`:
+1. Generates query embedding via `RubyLLM.embed`
+2. Finds nearest 8 `ContentChunk` records
+3. Builds system prompt with retrieved context + Ruby-only instructions
+4. Calls `ask(question, with_instructions: system_prompt, &block)`
+
+### Chat purpose
+
+Add `"rag"` to existing purposes (`conversation`, `summary`, `testimonial_generation`, `testimonial_validation`). `ChatResponseJob` detects `purpose == "rag"` and calls `ask_with_rag`.
+
+### Settings (no hardcoded models)
+
+Add `embedding_model` to `Setting::ALLOWED_KEYS`, default `"text-embedding-3-small"` (1536 dims). Configurable via the existing Madmin settings page, same pattern as `summary_model`, `testimonial_model`, etc.
+
+## Routes
+
+```ruby
+# Top-level, no team slug exposed
+resource :chat, only: [:show, :create], controller: "chats"
+```
+
+- `GET /chat` — shows user's current chat (creates one if none), requires auth
+- `POST /chat` — sends a message, requires auth
+- Homepage input form posts to `/chat`
+
+## Auth flow
+
+The existing `authenticate_user!` in `ApplicationController` already handles the redirect:
+
+```ruby
+def authenticate_user!
+  unless user_signed_in?
+    session[:return_to] = request.original_url if request.get?
+    redirect_to github_auth_with_return_path, alert: "Please sign in with GitHub to continue."
+  end
+end
+```
+
+For POST from the homepage input, store the pending question in the session before redirecting to GitHub OAuth, then replay it after sign-in.
+
+## Files to create/modify
+
+| Action | File |
+|--------|------|
+| **Gemfile** | +`sqlite-vec`, +`neighbor` |
+| **database.yml** | Load vec0 extension alongside existing uuid extension |
+| **Migration** | `create_content_chunks` + `create_content_embeddings` (vec0) |
+| **Model** | `app/models/content_chunk.rb` (with `has_neighbors`) |
+| **Concern** | `app/models/concerns/embeddable.rb` |
+| **Concern** | `app/models/concerns/chat/rag_augmentable.rb` |
+| **Model mods** | Include `Embeddable` in Post, Testimonial, Project |
+| **Model mod** | Include `Chat::RagAugmentable` in Chat |
+| **Setting** | Add `embedding_model` to `ALLOWED_KEYS` |
+| **Controller** | New top-level `/chat` controller (or reuse existing with new route) |
+| **Job** | `generate_embedding_job.rb` |
+| **Job** | `backfill_embeddings_job.rb` (one-time for existing content) |
+| **Job mod** | `chat_response_job.rb` — detect `purpose == "rag"`, call `ask_with_rag` |
+| **View mod** | `home/index.html.erb` — add input + "Chat" nav link |
+| **Routes** | `config/routes.rb` — add top-level `/chat` |
+| **i18n** | Locale files for new strings |
+
+## Design decisions
+
+- **No hardcoded AI models** — everything via `Setting.get(...)`, same pattern as existing AI features
+- **No new chat views** — reuse the existing beautiful templates from the remote template
+- **No team slug in URLs** — multi-tenancy is internal, users see clean `/chat`
+- **No chat list, no multi-chat UI** — single rotating chat keeps UX simple and costs bounded
+- **No vector DB** — sqlite-vec keeps infrastructure footprint zero
+- **Auth required** — connects chats to users for cost tracking via existing `Costable`
+
+## Migration path if corpus grows
+
+If content grows to tens of thousands of documents or semantic quality becomes insufficient:
+- Switch embedding model via the setting (e.g., to `text-embedding-3-large`, 3072 dims)
+- Add chunking (split long posts into overlapping windows) in `GenerateEmbeddingJob`
+- Everything else (controller, concern, UI) stays the same
+
+## Open questions for implementation
+
+- Auto-rotation threshold: 30 messages? 24h? Both?
+- Should old rotated chats be visible to admins in Madmin? (Probably yes for debugging)
+- Rate limit per user: e.g., 20 messages/hour to prevent abuse even among logged-in users
+- Do we index draft/unpublished content? (No — only published)
+- Should failed embeddings retry or silently skip? (Retry with backoff via SolidQueue defaults)
diff --git a/test/models/message_test.rb b/test/models/message_test.rb
index 89d440d..f4e1f78 100644
--- a/test/models/message_test.rb
+++ b/test/models/message_test.rb
@@ -1,6 +1,8 @@
 require "test_helper"
 
 class MessageTest < ActiveSupport::TestCase
+  include ActiveJob::TestHelper
+
   test "calculates cost based on token usage" do
     message = Message.new(
       chat: chats(:one),
@@ -15,4 +17,14 @@ class MessageTest < ActiveSupport::TestCase
 
     assert message.cost > 0
   end
+
+  test "broadcasts create using messages/message partial regardless of role" do
+    %w[user assistant system tool].each do |role|
+      message = Message.new(chat: chats(:one), role: role, content: "hi")
+
+      assert_nothing_raised do
+        perform_enqueued_jobs { message.save! }
+      end
+    end
+  end
 end