AI assistant for a 20-person corporate team. Ask any question about 20GB+ of project documents — Word, Excel, PDF, PowerPoint, source code. Answers via Microsoft Teams and Web UI. Gets smarter over time through The Enrich.
A Real Question
"@neo what do you mean by F1 DRS?" — from question to answer, with The Enrich
Everything Inside Cortex
Cortex is the sun — Smart RAG pipeline components radiate out
The Enrich — How Cortex Gets Smarter
Weak answer → Trinity reads source docs → writes deep explanation → next time the answer is great
"DRS is related to aerodynamic systems and rear wings. It is mentioned in the racing regulations document..."
😐 Thin. Vague. Useless. GPT-4.1 only got random chunks — a table header from a spreadsheet, half a sentence from a rules PDF. Not enough context to give a real answer.
⚡ ENRICH RUNS
📝 Step 1 — Spot the weakness
Cortex logs every question. Trinity scans the log and sees: "F1 DRS — asked 4 times, answer always thin"
🔺 Step 2 — Trinity deep-reads the source
Trinity (Opus 4.6) doesn't just read chunks — she reads the full documents:
📄 F1_Technical_Regulations_2026.pdf — all 80 pages
📊 DRS_Zone_Specifications.xlsx — every circuit
📽️ Active_Aero_2026_Changes.pptx — slide by slide
Then writes a complete, flowing explanation — like a senior engineer would explain to a new joiner.
💎 Step 3 — Load into ChromaDB + FTS
The enriched explanation is:
→ Chunked into proper sections → Embedded as vectors into ChromaDB (384-dim) → Indexed as keywords into SQLite FTS5 → Scored 1.2× higher than raw document chunks
Next time someone searches, enriched content surfaces FIRST.
✅ AFTER ENRICH
💬 MS Teams — Same question, next day
@neo what do you mean by F1 DRS?
💬 MS Teams — @neo replies
"DRS stands for Drag Reduction System — a movable rear wing flap introduced in 2011 to promote overtaking. When a driver is within 1 second of the car ahead at a DRS detection point, they can open the flap on the following straight, reducing drag by ~20% and gaining up to 12 km/h. Each circuit has 1-3 DRS zones. Under the 2026 regulations (DRS-2026), the traditional DRS flap is eliminated entirely — replaced by full active aerodynamics across the whole car, allowing continuous drag/downforce adjustment."
🎯 Detailed. Confident. Actually useful. Same GPT-4.1 model. Same question. But now it has enriched source material to work with — that's the difference.
↺ Every ENRICH run makes Cortex smarter · 20 enrichments written so far · Not retraining the AI — giving it better source material
✕
🔮 ChromaDB — Vector Search (Semantic)
What it does
ChromaDB stores 349,000 text chunks from 20GB of documents — but not as plain text. Each chunk is converted into a 384-dimension number vector (an "embedding").
These vectors capture meaning, not words. So when you search, it finds chunks that are about the same thing — even if they use completely different words.
The question: "@neo what do you mean by F1 DRS?"
ChromaDB searches 349K chunks for meaning. No doc says "F1 DRS" directly — but these are the closest matches:
📄 "The automobile engine burns fuel at 15,000 RPM..."
0.71
📄 "Aerodynamic wind tunnel testing for racing..."
0.65
📄 "Vehicle maintenance procedures..."
0.58
These are related to racing but none actually explain DRS. GPT-4.1 stitches them together and gets:
😐 Answer #1: "DRS is related to aerodynamic systems and rear wings..."
Garbled. Didn't answer the question.
⚡ Trinity enriches → uploads back to ChromaDB
🔮 Same question AFTER Enrich — new scores
Trinity wrote a deep explanation and loaded it into ChromaDB. Now:
FTS5 is Full-Text Search — the old-school way. It indexes every word in every document and finds exact matches.
Fast, precise, and catches things that semantic search misses — like part numbers, regulation codes, driver names, and technical specs.
Example: "DRS-2026" — only keywords work here
You search: "DRS-2026" (the drag reduction system regulation code)
📄 "...DRS-2026 eliminates the rear wing flap..."
✓ exact hit
📄 "...under DRS-2026, active aero replaces..."
✓ exact hit
📄 "...overtaking strategy and slipstream..."
✗ no match
Semantic search would miss "DRS-2026" entirely — it doesn't understand regulation codes. FTS5 finds it instantly.
Why BOTH are needed (Hybrid Search)
ChromaDB = "what does the question mean?" (F1 DRS → aerodynamic systems ✓) FTS5 = "what exact words appear?" (DRS-2026 → exact hit ✓)
Cortex runs both searches in parallel, merges results, and sends the best chunks to GPT-4.1. This is called Hybrid Search — you get the best of both worlds.
Also includes enriched docs (Trinity-written deep explanations) scored higher than raw chunks.