AI Code Review Assistant

An intelligent, conversational code review platform

Submit code in any language, receive structured AI-powered feedback with severity-rated issues, then ask follow-up questions in a live chat — all driven by Azure OpenAI and delivered in real time.

---

---

Highlights

Languages 12 languages supported Ruby, Python, JavaScript, TypeScript, Go, Rust, Java, C#, PHP, Swift, Kotlin, SQL Scored Reviews Quality score (1-10) with findings categorized across five dimensions: Bugs, Security, Performance, Code Quality, Best Practices

Live Chat Ask the AI to explain issues, show fixes, or suggest alternatives — full conversational context preserved Real-Time Updates No page refresh — everything streams live via WebSockets (Turbo Streams) Zero Extra Infrastructure PostgreSQL is the only dependency — no Redis, no Sidekiq, no separate message broker

Submit Code  ──>  AI Reviews It  ──>  Structured Feedback  ──>  Ask Follow-Up Questions
 (any language)      (GPT-4o)       (scored, categorized)        (conversational chat)

---

Tech Stack

Layer	Technology	Version	Role
Core
	Ruby on Rails	8.0.4	Full-stack web framework
	PostgreSQL	14+	Primary database
AI / LLM
	LangChainrb	0.19.5	AI orchestration framework
	Azure OpenAI	GPT-4o	LLM provider
	ruby-openai	8.3.0	OpenAI API client
Frontend & Real-Time
	Hotwire	—	Turbo + Stimulus (SPA-like, no JS framework)
	Turbo Streams	—	Real-time DOM updates over WebSockets
	esbuild	—	JavaScript bundling
	Propshaft	—	Asset pipeline
Infrastructure
	Solid Queue	—	Background job processing (DB-backed)
	Solid Cache	—	Database-backed caching
	Solid Cable	—	WebSocket pub/sub (Action Cable adapter)
	Puma	—	Web server
	Kamal	—	Docker-based deployment

---

Features in Detail

1. Structured AI Code Review

User submits code  ──>  Prompt Template + Few-Shot Examples  ──>  Azure OpenAI (GPT-4o)
                                                                         |
                    <────────────────────────────────────────────────────-┘
                    v
          StructuredOutputParser (JSON Schema Enforcement)
                    |
                    v
     ┌──────────────────────────────────────────┐
     |  { overall_score: 7,                     |
     |    summary: "Generally clean code...",   |
     |    bugs: [...],                          |
     |    security: [...],                      |
     |    performance: [...],                   |
     |    code_quality: [...],                  |
     |    best_practices: [...] }               |
     └──────────────────────────────────────────┘

• Each issue includes: severity (critical / major / minor), line number, description, suggestion
• Malformed JSON from the LLM is auto-corrected via OutputFixingParser
• Token usage (prompt + completion + total) tracked per review

2. Few-Shot Prompt Engineering

Language-specific YAML examples teach the AI what quality looks like:

app/prompts/examples/
├── ruby.yml         # N+1 queries, mass assignment, find_by vs where.first
├── python.yml       # SQL injection, hardcoded credentials, non-Pythonic patterns
├── javascript.yml   # XSS vulnerabilities, missing error handling, === vs ==
└── default.yml      # Generic fallback

Adding a new language or improving review quality requires zero code changes — edit or add a YAML file.

3. Conversation Memory (Follow-Up Chat)

You:   "Can you explain the SQL injection issue in more detail?"
AI:    "The SQL injection on line 15 occurs because user input is
        interpolated directly into the query string..."

You:   "Show me the corrected version."
AI:    "Here's the fixed code using parameterized queries: ..."

You:   "Would using a scope be even better?"
AI:    "Yes! Here's a scope-based approach: ..."

• Powered by Langchain::Assistant — manages message history automatically
• Conversation history persisted in messages table and reloaded for each request
• Initial code + review seeded as first two messages when a review completes
• Chat suggestions appear as clickable pills for common questions

4. Real-Time UI

The review page uses a two-column layout with live updates:

┌──────────────────────────────────────────────────────────────────────┐
|  Review Title                       [ruby] [Completed] [8/10]       |
├─────────────────────────────────┬────────────────────────────────────┤
|                                 |                                    |
|  Submitted Code                 |  Follow-Up Chat                    |
|  ┌───────────────────────────┐  |  ┌──────────────────────────────┐  |
|  | def find_user(id)         |  |  | Explain the critical         |  |
|  |   User.where(id: id)     |  |  | issue              [Send]    |  |
|  |       .first             |  |  └──────────────────────────────┘  |
|  | end                       |  |                                    |
|  └───────────────────────────┘  |  ┌──────────────────────────────┐  |
|                                 |  | [You] Explain the critical   |  |
|  AI Feedback                    |  |  issue                        |  |
|  ┌───────────────────────────┐  |  |                               |  |
|  | 7/10 - Generally clean.. |  |  | [AI] The main issue is on    |  |
|  |                           |  |  |  line 2 where...             |  |
|  | Bugs & Errors (1)         |  |  └──────────────────────────────┘  |
|  | | minor | find_by...     |  |                                    |
|  |                           |  |  ┌──────────┐  ┌──────────────┐   |
|  | Security (1)              |  |  | Show fix |  | Security ?   |   |
|  | ...                       |  |  └──────────┘  └──────────────┘   |
|  └───────────────────────────┘  |                                    |
|           col-8                 |             col-4                  |
├─────────────────────────────────┴────────────────────────────────────┤
|  [Back to Reviews]   [New Review]                                    |
└──────────────────────────────────────────────────────────────────────┘

• Left column (col-8): Submitted code + AI feedback with scored issue cards
• Right column (col-4): Sticky chat panel that stays visible while scrolling
• Stacks to single column on screens < 1024px

---

Architecture

┌──────────────────────────────────────────────────────────────────────┐
|                            Rails App                                 |
|                                                                      |
|  ┌───────────────────┐      ┌─────────────────────────────────────┐  |
|  | CodeReviews       |─────>| CodeReviewJob (Solid Queue)         |  |
|  | Controller        |      |                                     |  |
|  └───────────────────┘      |  ┌───────────────────────────────┐  |  |
|                              |  | CodeReviewService              |  |  |
|  ┌───────────────────┐      |  |  - Prompt Templates            |  |  |
|  | CodeReviewMessages |     |  |  - Few-Shot Examples            |  |  |
|  | Controller        |      |  |  - Structured Output Parser    |  |  |
|  └───────┬───────────┘      |  └──────────────┬────────────────┘  |  |
|          |                   └────────────────┬┘                   |  |
|          |                                    |                    |  |
|          v                                    v                    |  |
|  ┌───────────────────┐      ┌─────────────────────────────────┐   |  |
|  | ChatMessageJob    |      | Turbo::StreamsChannel            |   |  |
|  | (Solid Queue)     |      |  broadcast_update_to             |   |  |
|  |                   |      |   - #code_review_feedback        |   |  |
|  | ┌───────────────┐ |      |   - #header_meta                |   |  |
|  | | Assistant     | |      |   - #chat_panel                 |   |  |
|  | | Service       | |      └─────────────────────────────────┘   |  |
|  | | (LangChain    | |                                            |  |
|  | |  Assistant)   | |──────>  Azure OpenAI (GPT-4o)              |  |
|  | └───────────────┘ |                                            |  |
|  └───────────────────┘                                            |  |
└──────────────────────────────────────────────────────────────────────┘
          |                                  |
          └───────── PostgreSQL ─────────────┘
               (data + jobs + cache + cable)

Request Flow

Code Review Flow

1. User submits code via the form → CodeReviewsController#create
2. CodeReviewJob runs in the background (Solid Queue)
3. CodeReviewService builds prompts (system + few-shot + user), calls Azure OpenAI via LangChain
4. Response parsed into structured JSON (with auto-fix on malformed output)
5. Conversation seeded — initial code + review saved as messages for future chat
6. Three Turbo Streams broadcast live: feedback, header badges, and chat panel

Follow-Up Chat Flow

1. User types a question → CodeReviewMessagesController#create
2. User message saved to DB and broadcast instantly via Turbo Stream
3. Typing indicator (animated dots) appears in the chat
4. ChatMessageJob initializes CodeReviewAssistantService with full conversation history
5. Langchain::Assistant generates a context-aware response
6. Reply saved to DB, typing indicator removed, AI response broadcast to chat

---

Database Schema

code_reviews

Column	Type	Description
title	string	Review title (required)
language	string	Programming language (default: ruby)
code	text	Submitted code snippet (required)
feedback	text	Raw AI feedback text
status	integer	0 pending → 1 processing → 2 completed → 3 failed
structured_feedback	jsonb	Parsed JSON: score, summary, categorized issues
overall_score	integer	Code quality score (1-10)
prompt_tokens	integer	LLM prompt token count
completion_tokens	integer	LLM completion token count
total_tokens	integer	Total token usage

messages

Column	Type	Description
code_review_id	bigint	FK to code_reviews
role	string	user, assistant, or system
content	text	Message body
prompt_tokens	integer	Token usage (if tracked)
completion_tokens	integer	Token usage (if tracked)
total_tokens	integer	Token usage (if tracked)

---

Project Structure

app/
├── assets/stylesheets/
│   └── application.css                    # Custom CSS: grid, cards, chat, responsive
├── controllers/
│   ├── code_reviews_controller.rb         # index, show, new, create, retry
│   └── code_review_messages_controller.rb # Chat message creation + Turbo broadcast
├── javascript/controllers/
│   └── chat_controller.js                 # Stimulus: send, suggestions, auto-resize, scroll
├── jobs/
│   ├── code_review_job.rb                 # Background review + 3 Turbo Stream broadcasts
│   └── chat_message_job.rb               # Background chat response + Turbo Stream
├── models/
│   ├── code_review.rb                     # Status enum, has_many :messages
│   └── message.rb                         # belongs_to :code_review, role validation
├── prompts/
│   ├── code_review_prompt.rb              # Templates, JSON schema, few-shot builder
│   └── examples/                          # Language-specific YAML few-shot examples
│       ├── ruby.yml
│       ├── python.yml
│       ├── javascript.yml
│       └── default.yml
├── services/
│   ├── code_review_service.rb             # Review orchestration: prompt → LLM → parse → save
│   └── code_review_assistant_service.rb   # Chat: LangChain Assistant + conversation memory
└── views/
    ├── code_reviews/
    │   ├── show.html.erb                  # Two-column layout (col-8 + col-4)
    │   ├── _feedback.html.erb             # Score bar + categorized issue cards
    │   ├── _chat.html.erb                 # Chat UI with Stimulus + Turbo Streams
    │   ├── _chat_panel.html.erb           # Conditional: chat / waiting / failed state
    │   └── _header_meta.html.erb          # Status + score badges (broadcast target)
    └── code_review_messages/
        └── _message.html.erb              # Chat bubble (user right / AI left)

config/
├── initializers/langchain.rb              # Azure OpenAI LLM setup (LANGCHAIN_LLM constant)
└── routes.rb                              # RESTful routes + nested messages

---

API Routes

Method	Path	Action
GET	/	List all reviews
GET	/code_reviews/new	New review form
POST	/code_reviews	Submit code for review
GET	/code_reviews/:id	View review + chat
POST	/code_reviews/:id/retry	Retry a failed review
POST	/code_reviews/:id/messages	Send a follow-up chat message

---

Getting Started

Prerequisites

Requirement	Version
Ruby	3.3.1+
Node.js	20.x
Yarn	1.x or 4.x
PostgreSQL	14+

1. Clone & Install

git clone https://github.com/mahendraxceed/ai_code_review_assistant.git
cd ai_code_review_assistant
bin/setup

2. Configure Azure OpenAI

Create a .env file in the project root:

AZURE_OPENAI_API_KEY=your-api-key
AZURE_OPENAI_ENDPOINT=https://your-resource.openai.azure.com
AZURE_OPENAI_DEPLOYMENT=your-deployment-name
AZURE_OPENAI_API_VERSION=2024-02-15-preview

Note: .env is gitignored — never commit API keys.

3. Run

bin/dev

Visit http://localhost:3000 — submit some code and watch the AI review it in real time.

---

Design Decisions

Decision	Rationale
PostgreSQL only	Solid Queue, Solid Cache, and Solid Cable all use Postgres — zero external infrastructure beyond the database
Custom CSS grid	Lightweight .row / .col-8 / .col-4 system. Wide 1400px container for reviews, 960px for everything else. No framework bloat
Background jobs for LLM	AI calls take 3-15s. Solid Queue keeps the request cycle fast; Turbo Streams deliver results the moment they're ready
Conversation seeding	When a review completes, initial code + AI feedback are saved as the first two messages. The LangChain Assistant gets full context for follow-ups
Partial-based broadcasts	Each live section has its own partial and target ID. broadcast_update_to swaps inner HTML while preserving the wrapper
YAML few-shot examples	Review quality is improved by editing YAML files, not code. Easy to add new languages or refine existing examples

---

Testing

bin/rails test          # Unit and integration tests
bin/rails test:system   # System tests (requires Chrome)

---

License

This project is proprietary. All rights reserved.