Hey friends 
I started by just helping a few devs troubleshoot weird RAG issues on forums.
Chunking bugs. Missing context. FAISS pulling totally unrelated documents.
At first, I thought: “Easy fix. Probably bad chunking or wrong embedding.”
But then I saw it happen again. And again. And again.
Not just random edge cases — core architectural flaws.
So instead of replying one by one, I wrote a full semantic theory + open-source tool to fix it.
Here’s the fire, and here’s my extinguisher:
What’s Actually Breaking in RAG
1. Token-based chunking kills meaning.
Chunking by character or token count sounds good — until your sentence gets sliced mid-thought and the embedding captures noise instead of intent.
2. Retrieval is semantically blind.
Cosine similarity ≠ semantic relevance. You get matches that look similar vector-wise, but are irrelevant in context.
3. System prompts can’t fight bad retrieval.
Once junk gets injected into the LLM context, it’s too late. You hallucinate confidently… with citations.
My Solution: WFGY Engine (Semantic Firewall)
I built the WFGY architecture — think of it as a semantic meaning firewall for your RAG pipeline.
Components:
ΔS(Delta Semantic Drift): Quantifies how much semantic distortion a chunk has from the original query.λ_observe: Measures the internal meaning-focus of retrieved chunks. Think: coherence entropy.- Semantic-Aware Chunking: Splits not by token count but by concept boundary.
- Layered Retrieval Filtering: Filters embeddings based on ΔS thresholds before hitting the LLM.
- Prompt Firewall Injection: Dynamically rewrites prompt instructions to guard hallucination-prone flows.
Visual Architecture (Click to see full PDF):
WFGY PDF: Full Paper + Math
Endorsed by Tesseract.js creator (36k GitHub stars)
2,000+ downloads in 1 month
Example Use Case:
Before:
“What’s the warranty policy on X?”
→ Retrieves FAQ chunk with unrelated refund info
→ Hallucinated answer: “X has a 7-day warranty with cash payout option”
After WFGY:
Same question → ΔS rejects refund chunk
→ Only warranty paragraph passes
→ Accurate response with fallback confidence score
Closing Drunk Thoughts
I thought this was going to be a one-off “just fix your chunk size” thing.
Turns out it’s a semantic alignment problem across the entire pipeline.
If you’re building with RAG and feel like you’re duct-taping your way out of a tornado — yeah, I feel you. Let’s stop hallucinating and start firewalling. Drop a comment if you’re in the same hole.
And if your LLM still hallucinates after this, maybe it’s just drunk.
