QYBER LLM Council

QYBER LLM Council is a full‑stack AI assistant that combines multiple LLMs into a 3‑stage council (Stage 1/2/3), with:

• real‑time chat interface
• document upload (PDF, DOCX, PPT/PPTX)
• conversation management
• export of the final answer to DOCX
• dark UI inspired by ChatGPT

The project was developed in pair‑programming with an AI assistant.

---

Features

Multi‑turn conversations

• Create and select conversations in the sidebar.
• The LLM remembers context: previous messages are included in the prompt, with a character limit to stay within model context size.
• You can delete conversations from the sidebar.
• Conversation titles are generated automatically from the first user message.

3‑Stage LLM Council

For each user message, the backend runs a 3‑stage pipeline:

1. Stage 1 – Individual Responses
   Several models (COUNCIL_MODELS) answer the question independently.
   The frontend shows these answers in a tab view (one tab per model).

2. Stage 2 – Peer Rankings
   Each model sees anonymized responses from all models (Response A, B, C, …) and:

   • analyzes them,
   • produces a ranking.

   The frontend displays:

   • a tab per evaluating model,
   • an “Aggregate Rankings (Street Cred)” section that combines all votes.

3. Stage 3 – Final Council Answer
   A “Chairman” model (CHAIRMAN_MODEL) synthesizes everything into a single final answer, shown in a dedicated dark‑themed block.

Models (Council members)

Models are configured in backend/config.py, for example:

• openai/gpt-5.1
• google/gemini-3-pro-preview
• anthropic/claude-sonnet-4.5
• x-ai/grok-4
• perplexity/sonar-pro-search

The Chairman is one of these models (by default anthropic/claude-sonnet-4.5).

Document upload & context

• Upload PDF, DOCX, PPT/PPTX via:
  • the + button in the input bar, or
  • drag & drop directly onto the input bar.
• Text extraction on the backend using pypdf, python-docx, python-pptx (see backend/file_extractor.py).
• Extracted text is added to the conversation as a special user message:
  • stored in the conversation JSON files,
  • included in the context sent to the models.
• On the frontend:
  • uploaded documents appear as chips / mini cards in the input area,
  • in history they appear as small “📎 filename” pills instead of dumping all the text.

DOCX export

• A backend endpoint generates a .docx file from the Stage 3 final answer.
• On the frontend, the last assistant message that has stage3 shows an “Export DOCX” button:
  • click → calls the export API,
  • triggers download of conversation-XXXX-final.docx.

User interface

• Dark theme:
  • Sidebar: #181818
  • Main chat area: #212121
  • Input bar: #303030 (pill shape)
  • Send button: round blue button.
• Sidebar:
  • centered QYBER logo
  • + New Conversation button
  • “Projects” section (static list)
  • “Your conversations” section (actual conversations).
• Chat:
  • empty state with “Start a conversation / Ask a question to consult QYBER”
  • user messages in dark bubbles
  • Stage 1/2/3 blocks using the same dark theme.

---

Architecture

Backend

• Framework: FastAPI (Python 3.10+)
• Main modules:
  • backend/main.py – HTTP endpoints:
    • /api/conversations (list/create)
    • /api/conversations/{id} (get/delete)
    • /api/conversations/{id}/message (non‑streaming council run)
    • /api/conversations/{id}/message/stream (SSE, streaming Stage 1/2/3)
    • /api/conversations/{id}/upload (upload document + text extraction)
    • /api/conversations/{id}/export-docx (generate and download DOCX)
  • backend/council.py – 3‑stage council logic, aggregate rankings, title generation
  • backend/openrouter.py – wrapper around the OpenRouter Chat Completions API
  • backend/storage.py – conversation storage as JSON under data/conversations/
  • backend/file_extractor.py – PDF/DOCX/PPTX text extraction
  • backend/config.py – configuration (API key, models, data directory)
• Context management:
  • conversation history is concatenated into a text history string,
  • truncated to MAX_HISTORY_CHARS to avoid exceeding model context.

Frontend

• Framework: React + Vite
• Libraries:
  • react-markdown for rendering markdown answers
  • native EventSource for SSE streaming from the backend.
• Key components:
  • App.jsx – top‑level layout, selected conversation, loading state
  • Sidebar.jsx / Sidebar.css – logo, “Projects”, “Your conversations”, new conversation and delete
  • ChatInterface.jsx / ChatInterface.css – messages, Stages, new input bar (pill), drag‑and‑drop, attachment chips, Export DOCX button
  • FileUpload.jsx – document upload logic and integration with the upload API
  • Stage1.jsx / Stage1.css – individual responses view
  • Stage2.jsx / Stage2.css – peer rankings + aggregate ranking
  • Stage3.jsx / Stage3.css – final council answer view
  • api.js – small API client (conversations, messages, streaming, upload, export DOCX).

---

Installation & Setup

Prerequisites

• Python 3.10+
• uv for Python dependency management
• Node.js + npm
• An OpenRouter account + API key

1. Configure environment

Create a .env file at the project root:

OPENROUTER_API_KEY=your_openrouter_api_key