Neocortex

Less Tokens. More Answers.

Comment

Neocortex

Inspiration

Modern AI coding assistants struggle with unfamiliar codebases. They dump entire files into context or miss critical dependencies. We asked: what if an AI could understand code the way a senior engineer does - by following function calls, understanding relationships, and retrieving only what's relevant?

Neocortex was born from watching LLMs hallucinate about code they couldn't see while ignoring the exact functions that mattered.

What it does

Neocortex is a semantic code retrieval engine that gives AI laser-focused context. Point it at any GitHub repository, ask a question like "how does authentication work?", and it returns the exact functions, classes, variables, and types needed to answer, compressed to fit your token budget.

Three stages:

  1. Parse - Tree-sitter extracts every function, class, and type into a searchable dependency graph
  2. Search - OpenAI embeddings + keyword matching + graph traversal finds relevant code
  3. Compress - Intelligent slicing maximizes relevance within your token budget

Instead of 17,000 tokens of code, you get 1,200 tokens of exactly what matters.

We integrated this as a full MCP (Model Context Protocol) server so any AI assistant can use Neocortex as a tool.

MCP Tools:

Tool Description
index_repo Clone and index any GitHub repository
search_code Semantic search with natural language
resolve_symbol Find definitions by exact name
get_snippet Get code from specific file:lines
classify_query Analyze query intent for search optimization

How we built it

Pipeline architecture in TypeScript:

  1. AST Parsing - Tree-sitter extracts code entities and relationships
  2. Dependency Graph - Bidirectional edges tracking function calls
  3. Vector Embeddings - OpenAI text-embedding-3-small for semantic search
  4. Query Classifier - Tunes search strategy based on query type
  5. Compression Pipeline - Query-aware slicing to cut tokens while preserving relevance
  6. MCP Server - LeanMCP exposes everything as AI-callable tools

Tree-sitter was chosen because it powers GitHub's syntax highlighting and provides precise AST access. The hybrid search (embeddings + keywords + graph expansion) outperforms pure semantic search on code.

Challenges we ran into

Embedding score calibration: OpenAI embeddings return 0.2-0.4 scores for relevant code, not 0.8+ like expected. We spent hours thinking search was broken before recalibrating thresholds.

Graph explosion: Expanding dependencies 2 levels deep can return hundreds of entities. We implemented depth limiting and prioritization to keep results focused.

Token budgets: Fitting meaningful code into 2,000 tokens while preserving context required building a slicing algorithm that keeps signatures, relevant lines, and truncates intelligently.

Sleep deprivation: At 3 AM, every bug looks like a feature. Caffeine got us through.

Accomplishments that we're proud of

Our benchmarks across 4 repositories (56 queries):

Metric Result
Average token reduction 85.2%
Cost savings 93.8%
Search latency 246ms
Debugging queries 95.6% reduction
Multi-hop queries 95.5% reduction

By Token Budget:

Budget Reduction
500 tokens 93.9%
1,000 tokens 87.0%
2,000 tokens 74.5%
4,000 tokens 86.1%

Cost Savings (GPT-4o):

Scale Monthly Savings
10,000 queries $808
100,000 queries $8,080

Other wins:

  • Hybrid search beats pure semantic search
  • Query classification auto-adjusts strategy
  • Full MCP server with 5 tools
  • Zero-config for any TypeScript/JavaScript repo
  • Team collaboration integrated seamlessly

What we learned

  • Embedding similarity scores vary wildly between models - calibrate empirically
  • Graph-based context expansion is crucial - functions don't exist in isolation
  • Token budgets force information density thinking
  • Tree-sitter is powerful once you learn its query patterns
  • MCP is the future of AI tool integration
  • "Power naps" during hackathons are a gamble

What's next for Neocortex

  • Multi-language support (Python, Go, Rust)
  • VS Code extension
  • SDK
  • Cached indexes for instant queries
  • Fine-tuned reranker for code relevance
  • MCP registry publishing

Built With

  • TypeScript
  • Tree-sitter
  • OpenAI Embeddings
  • LeanMCP
  • Node.js
  • Express.js
  • React + Vite
  • Tailwind CSS

Try It

git clone https://github.com/rusheelsharma/neocortex
cd neocortex && pnpm install

# Run demo
pnpm demo

# Query any repo
pnpm dev context https://github.com/user/repo "how does X work?"

# Start MCP server
pnpm mcp:dev

Built by Software Developers for Software Developers 💪

Built With

Share this project:

Updates