⬡ Under The Hood

How the AI Engine Works

A technical walkthrough of every step — from natural language input to a contextually ranked list of spaces. Built on semantic embeddings, graph-based orchestration, and multi-factor scoring.

The Big Picture

When you submit an event request, your text travels through a 5-node AI pipeline built with LangGraph. Each node does one job: extract intent, apply constraints, match semantically, score dynamically, and generate an explanation. The whole thing runs in under 2 seconds.

Node 1 of 5

Intent Extraction

The engine reads your plain-English description and extracts structured intent: What type of event is this? What atmosphere do you want? How important is luxury? How much noise is acceptable? Is this social or professional?

This is done by a language model prompt that outputs a JSON object. For example, "outdoor rooftop wedding for VIP guests at sunset" becomes:

# Input (your text)
"Outdoor rooftop wedding reception for 100 guests, ultra-luxury VIP experience at sunset"

# Output (structured intent JSON)
{
  "event_type": "wedding",
  "atmosphere": "romantic",
  "luxury_expectation": 0.92,
  "noise_tolerance": 0.5,
  "social_context": "celebration",
  "indoor_preference": false
}
      

LangGraph Node Prompt Engineering JSON Schema Output

Node 2 of 5

Constraint Analysis

Now the engine applies hard filters — rules that eliminate spaces that simply can't work. If you need 100 guests, a 22-seat boardroom is out, regardless of how good the AI match score is.

Constraints checked include: guest count vs. capacity, indoor/outdoor preference, availability status, and weather compatibility (e.g., outdoor spaces get flagged on rainy days).

Hard Filters Weather Logic Capacity Check

Node 3 of 5

Semantic Matching via Embeddings

This is the core of the ML system. Each space in the database has been pre-converted into a 384-dimensional vector (called an embedding) that represents its meaning. Your query is also converted into a vector — then we find the spaces with the most similar vectors.

Think of it like converting words into coordinates in 384-dimensional space. "Rooftop wedding" lands close to "Infinity Pool Deck" and "Skyline Rooftop" because their descriptions share semantic meaning, even if the exact words differ.

Cosine Similarity Formula

similarity(query, space) = (query · space) / (|query| × |space|)

Range: 0.0 (no match) → 1.0 (perfect match)

# Your description → embedding vector (384 numbers)
query_vec = model.encode("outdoor rooftop wedding...")
# [0.021, -0.143, 0.087, 0.312, ... 384 values total]

# Compare against all space vectors using FAISS
scores = faiss_index.search(query_vec, top_k=20)
# Returns top 20 most similar spaces instantly
      

Example embedding vector (first 32 of 384 dimensions):

Each color cell = one number. Blue = positive, red = negative, intensity = magnitude.

sentence-transformers all-MiniLM-L6-v2 model FAISS vector index 384 dimensions

Node 4 of 5

Dynamic Contextual Scoring

Semantic similarity alone isn't enough. A ballroom might match "elegant wedding" well semantically, but if you need 300 guests and it only fits 80, that's a problem.

Node 4 blends the semantic score with contextual factors — luxury level, noise tolerance, weather suitability, and capacity fit — into a single Final Score. It also detects whether a space has a SpaceTransformation available (e.g., Pool Deck → Covered Cocktail Reception on rainy days) and boosts its score accordingly.

Final Score Formula

final = (semantic × 0.50) + (luxury_fit × 0.20) + (capacity_fit × 0.15) + (noise_fit × 0.10) + (weather_fit × 0.05)

+ 0.15 bonus if transformation available on rainy/windy day

Semantic

50%

How well the description matches the space's meaning

Luxury Fit

20%

Does the space's luxury level match your VIP expectation?

Capacity Fit

15%

How well guest count fits (penalty if over or under)

Noise Fit

10%

Match between space noise level and event type

Weather Fit

Is the space suitable for the reported weather?

Weighted Scoring SpaceTransformation detection Weather boost logic

Node 5 of 5

Explanation Generation

The final node generates a plain-English explanation for why each space was recommended. It references the specific factors that drove the score: semantic match, luxury level, capacity fit, and any active transformation.

If weather is rainy or windy and the space has a transformation available, the explanation specifically calls out the conversion option and how long setup takes.

# Example output for Skyline Rooftop on a rainy day
"Skyline Rooftop scores highly for an outdoor wedding with panoramic views
and a luxury score of 10/10. Note: current weather is rainy — this space
can be converted to Enclosed Sky Pavilion in 240 min (8 staff required),
maintaining full skyline views with complete weather protection."
      

Template-based NLG Transformation-aware Per-space explanation

Why Not Just Use ChatGPT?

Capability	Generic LLM (ChatGPT)	This Engine
Knows your actual spaces	✗ No — hallucinates venue details	✓ Yes — reads live DB data
Real-time availability check	✗ No	✓ Yes — queries availability_status
Transformation logic	✗ No	✓ Yes — weather-aware boost + badge
Reproducible scoring	~ Inconsistent — varies per run	✓ Deterministic — same input = same score
Works offline / no API cost	✗ Requires OpenAI API	✓ Runs locally — all models on-device
Auditable reasoning	~ Black box	✓ Every score is inspectable

Technology Stack

🐍

Django 4.2

v4.2.13

Web framework, ORM, admin panel, and URL routing

🔗

LangGraph

v0.2.28

5-node stateful AI workflow — orchestrates every step of the pipeline

🧠

sentence-transformers

all-MiniLM-L6-v2

Converts text descriptions into 384-dimensional semantic vectors

⚡

FAISS

faiss-cpu

Facebook AI Similarity Search — instant nearest-neighbor lookup over all space embeddings

🗺

Leaflet.js

v1.9.4

Interactive map showing building locations with CartoDB dark tiles

🗄

SQLite

Built-in

Local database storing buildings, spaces, transformations, and recommendation history