Back to Projects

TheStatic Knowledge

RAG System for Codebase Intelligence

Production
PythonFastAPILangChainChromaDBOllamanomic-embed-text
42K+
Chunks Indexed
Across 5 repos
97/s
Embedding Speed
GPU accelerated
<5s
Query Response
End to end
768
Vector Dimensions
nomic embeddings
RAG chat interface with query and response

Chat Interface

RAG dashboard overview

Dashboard

What It Does

TheStatic Knowledge is a production RAG (Retrieval-Augmented Generation) system that indexes the entire TheStatic ecosystem codebase. Ask any question about the architecture, API endpoints, smart contracts, or SDK - get accurate answers with source file references in seconds.

Built for my own development workflow, it eliminates context-switching between repositories and documentation. Query patterns span multiple repos seamlessly, like "How does a stream get from the backend to a Decentraland scene?"

Indexed Repositories

thestatic-tvNext.js frontend + FastAPI backend
~200 files
thestatic-dcl-sdkTypeScript SDK for Decentraland
~50 files
thestatic-contractsSolidity smart contracts
~20 files
thestatic-hqMain Decentraland venue
~30 files
Other DCL scenespopup, basic, starter scenes
~100 files

Architecture Highlights

Hybrid Search Pipeline

Two-stage retrieval: vector similarity search (top 10) followed by BM25 keyword scoring and weighted reranking (0.7 vector + 0.3 BM25) to get final 5 results.

GPU-Accelerated Embeddings

Ollama with nomic-embed-text model running on RTX 5090. Batch processing at 500 chunks per request achieves 97 chunks/sec sustained throughput.

Real-Time Index Management

Background indexing with live progress tracking. Indexes entire codebase in ~7 minutes. Persistent ChromaDB storage (356MB) for instant startup.

Clean LLM Output

Optimized prompts prevent instruction echoing. Stop token handling and post-processing ensure professional responses without artifacts.

Technical Details

  • Indexes 5 repositories: thestatic-tv, thestatic-dcl-sdk, thestatic-contracts, DCL scenes (hq, popup, basic, starter)
  • Recursive character text splitting with code-aware separators (class, function, export)
  • Metadata-rich chunks with source path, repo name, file type, and line numbers
  • FastAPI backend with async endpoints for chat, indexing, and progress tracking
  • Web dashboard with real-time chat interface and indexing progress visualization
  • BM25 scoring using term frequency, inverse document frequency, and length normalization

RAG Pipeline

1. LOAD - Read files from 5 repositories (.py, .ts, .sol, .md, .json)

2. SPLIT - Chunk with code-aware separators (500-1000 chars, 100 overlap)

3. EMBED - Generate 768-dim vectors via Ollama (nomic-embed-text)

4. STORE - Persist to ChromaDB with metadata (source, repo, type)

5. QUERY - Embed question, vector search (top 10)

6. RERANK - BM25 keyword scoring, weighted combine, top 5

7. GENERATE - LLM with context, clean response with sources

Example Query

"Explain the complete flow of how a live stream gets from the backend to a Decentraland scene"

The RAG system retrieves context from:

  • thestatic-tv/ - GCP Livestream API, RTMP ingest, HLS encoding
  • thestatic-dcl-sdk/ - SDK video player component, scene integration
  • thestatic-hq/ - How the venue uses the SDK

Response synthesizes the complete pipeline: broadcaster -> RTMP -> GCP -> HLS -> CDN -> SDK -> scene display