✦ Context Galaxy

A semantic memory workspace where conversations evolve into explorable galaxies

Context Galaxy is not a chatbot. It's a graph-based AI memory system where topics crystallize into planets, conversations become galaxies, and your AI learns you over time.

---

What Makes It Different

Most AI tools pass raw chat history into the LLM context window. Context Galaxy builds a semantic property graph from your conversations instead only storing what's user-specific and injecting only what's relevant. The LLM already knows general facts; this system only tracks your learning path, priorities, and evolving interests.

Standard AI             Context Galaxy
──────────────          ──────────────────────────────────
[msg 1]                 Planet: Learning LangChain  (root)
[msg 2]   →  LLM           └── Moon: RAG           (crystallized after 3 mentions)
[msg 3]                    └── Moon: Agents         (forming — 2/3)
[msg N]                    └── Archived: LangChain v0.1  (deep space)

---

Core Features

🪐 Galaxy Visualization

Built on React Flow with custom circular planet nodes, roots are large glowing violet spheres, branches are smaller blue moons, and candidates are dim proto-moons. Nodes are laid out by an orbital algorithm: root at cente with candidate nodes & branches at radius.

🧠 Graph-Based Semantic Memory

Memory is stored as a property graph in PostgreSQL via SQLAlchemy async ORM. Each node carries a label, type (root / branch / candidate), priority, score, embedding vector, and is_active flag. Edges store the cosine similarity weight between connected concepts. Three isolated layers: raw message log → dynamic context graph → visual galaxy.

🌱 Topic Crystallization Engine

New topics start as Candidates — they don't enter memory until they've been meaningfully discussed by user frequently . The system tracks mention_count per candidate. At 3 mentions (raised to 5 in long chats), the candidate topics crystallizes into a full graph node, gets connected to the root planet with a weighted edge, and disappears from the forming moons list. This prevents graph explosion from casual one-off mentions.

🔍 Semantic Deduplication

Before tracking any new candidate, two checks run in sequence:

1. Fuzzy string match : normalized comparison + substring containment catches "RAG" vs "RAG pipeline", and Levenshtein distance ≤ 2 catches typos like "retreival". If matched, increments the existing candidate instead of forking.

2. Embedding cosine similarity — sentence-transformers generates a 384-dim vector for the new topic and compares it against all existing candidates. Similarity ≥ 0.82 treats them as the same concept.

📡 Dual-Source Intent Extraction

Topics are extracted from both the user message and the assistant response (first 600 chars). This matters because casual user questions like "Prime minister tough job right?" have low noun density, but the assistant response contains the rich vocabulary ("Prime Minister", "policy making", "leadership"). Pipeline: spaCy for noun phrase chunking and NER → Groq LLM for semantic abstraction → structured JSON output.

📥 Context Jar Sidebar

Real time sidebar showing the live state of memory. Active Core tab: Shows crystallized nodes that are actively shaping LLM responses. Deep Space tab: shows archived older nodes. Forming Moons: Shows top5 candidates by mention count with a live progress bar that can turn into a subtopic node with active polling. Per node controls: Priority (HIGH / MEDIUM / LOW), activate/deactivate, delete.

💬 Streaming Chat with Adaptive Length

Chat streams via FastAPI SSE with httpx async streaming to OpenRouter on the backend. Before each request, max_tokens is set based on question type: 120 for yes/no confirmations, up to 1000 for deep dives using regex pattern classification. Also includes inline message editing with per message copy button, and a regenerate option.

🗂️ Chat Management

Left sidebar with a complete chat history inspired by modern day LLMs. With options for renaming and cascade deletion .

---

Tech Stack

Layer	Technology	Purpose
Frontend framework	Next.js 14 (TypeScript)	App routing, SSR, file-based pages
Graph visualization	React Flow	Custom planet nodes, orbital layout, minimap
Animations	Framer Motion	Sidebar transitions, node entrances
Backend API	FastAPI	Async REST endpoints + SSE streaming
ORM	SQLAlchemy 2.0 async	Async database queries with asyncpg
Schema validation	Pydantic v2	Request/response models, settings
Database	PostgreSQL	Property graph storage — nodes, edges, candidates
LLM orchestration	LangChain + LangGraph	Prompt management, stateful workflow
LLM gateway	OpenRouter	Routes to Claude, GPT-4o, Mixtral, etc.
Fast inference	Groq	Low-latency intent extraction
NLP preprocessing	spaCy	Tokenization, NER, noun phrase chunking
Embeddings	sentence-transformers	all-MiniLM-L6-v2, 384-dim vectors
Async HTTP	httpx	Streaming LLM calls to OpenRouter
Styling	Tailwind CSS	Deep space dark theme, utility-first

---

Project Structure

context-galaxy/
├── backend/
│   └── app/
│       ├── api/
│       │   └── chat.py              # REST + SSE endpoints, graph/context routes
│       ├── core/
│       │   ├── config.py            # Pydantic BaseSettings
│       │   └── llm.py               # OpenRouter / Groq client
│       ├── database/
│       │   └── session.py           # Async SQLAlchemy session
│       ├── models/                  # ORM models
│       │   ├── chat.py
│       │   ├── message.py
│       │   ├── context_node.py
│       │   ├── candidate_topic.py
│       │   └── context_edge.py
│       ├── nlp/
│       │   └── intent_extractor.py  # spaCy + Groq hybrid pipeline
│       ├── services/
│       │   └── chat_service.py      # stream_llm, crystallization, deduplication,
│       │                            # context prompt builder, pruning
│       └── main.py
│
└── frontend/
    └── src/
        ├── app/
        │   ├── page.tsx             # New galaxy home
        │   ├── chat/[id]/page.tsx   # Streaming chat thread
        │   └── galaxy/[id]/page.tsx # Galaxy Map (dynamic import, SSR-safe)
        └── components/
            ├── GalaxyVisualizer.tsx # React Flow — module-level nodeTypes (no infinite loop)
            ├── ContextJarSidebar.tsx
            ├── LeftSidebar.tsx
            └── MessageBubble.tsx    # Copy, edit, regenerate actions

---

Getting Started

Prerequisites

• Node.js 18+, Python 3.11+, PostgreSQL 15+
• OpenRouter API key

Backend

cd backend
python -m venv venv && source venv/bin/activate
pip install -r requirements.txt
python -m spacy download en_core_web_sm

# configure .env then:
python -m app.database.init_db
uvicorn app.main:app --reload --port 8000

Frontend

cd frontend
npm install
echo "NEXT_PUBLIC_API_URL=http://localhost:8000" > .env.local
npm run dev

Environment Variables

# backend/.env
DATABASE_URL=postgresql+asyncpg://user:password@localhost:5432/context_galaxy
OPENROUTER_API_KEY=sk-or-...
LLM_MODEL=anthropic/claude-3-haiku
GROQ_API_KEY=gsk_...

---

Database Schema (simplified)

chats          → id, title, created_at
messages       → id, chat_id, role, content, created_at
context_nodes  → id, chat_id, label, node_type, priority, score,
                 embedding[], is_active, mention_count, last_accessed
context_edges  → id, chat_id, source_id, target_id, weight
candidate_topics → id, chat_id, label, mention_count, threshold, last_seen

---

Future Roadmap

Status	Feature
🔄	LangGraph stateful multi-node workflow
🔄	pgvector for native embedding similarity queries
🔄	Memory decay scoring + auto-archive
📋	User authentication
📋	Cross-chat galaxy relationship map
📋	Export galaxy as JSON / GraphML

---

_{Built on the belief that AI memory should reflect you — not just your last message.}

✦ Context Galaxy