Download Latest Version
research.zip (194.2 MB)
Get an email when there's a new version of Research From AI
# WGSL Screen‑Driven Compute — RAG Seed Pack
This pack contains 12 WGSL examples (with metadata) ready to embed into AnythingLLM backed by LanceDB.
## Files
- `wgsl_rag_seed.jsonl` — one JSON object per line with fields:
- id, slug, title, category, difficulty, tags, summary, body_markdown, wgsl_code, io, dependencies
## Suggested Workspace Setup (AnythingLLM + LanceDB)
1. Create a new **Workspace** named “GVPIE WGSL RAG”.
2. Add **Data Source → Files** and upload `wgsl_rag_seed.jsonl`.
3. Ingestion settings (recommended):
- Chunking: **By document** (disable aggressive chunking for code)
- Max tokens per chunk: **1500–2000**
- Overlap: **0–50 tokens**
- Metadata extraction: **enabled** (title, category, tags, difficulty)
4. After embedding, try queries like:
- “r32float storage texture identity node”
- “convolution kernel from row strip”
- “temporal EMA previous frame feedback”
- “screen‑driven ring buffer queue”
5. Retrieval prompts (pattern):
- *“Give me the WGSL code only for {topic}, include I/O bindings.”*
- *“Summarize the protocol and return a short explanation plus the code fence.”*
## Tips
- Prefer **semantic** search over keyword-only.
- Use the `tags` and `category` metadata as filters to narrow recall.
- When generating new examples, re‑use the same metadata schema so future embeddings remain consistent.
Happy hacking!
