🔬 Technical Deep Dive

How SecPhi Actually Works

A transparent look at our architecture, what AI can and can't do, and honest answers about multi-agent systems.

🧠

What LLMs Can (and Can't) Do

Large Language Models (LLMs) like GPT-4, Claude, and Gemini are text-in, text-out machines. They receive text, process it, and generate text back. That's it. They cannot:

🌐❌ Cannot Access the internet

📡❌ Cannot Make API requests

📁❌ Cannot Read files directly

🗄️❌ Cannot Query databases

💡Key Insight

LLMs only "see" what your application code feeds them. When SecPhi shows CVE analysis, the AI didn't fetch that data—our backend code did, then packaged it into a prompt for the AI to analyze.

🔄

What Actually Happens

// SecPhi Backend (JavaScript/Node.js)

// Step 1: Our code makes API call

const response = await fetch('https://nvd.nist.gov/api/...')

const cveData = await response.json()

// Step 2: Our code builds a prompt WITH that data

const prompt = `

Here is real CVE data from NVD:

${JSON.stringify(cveData)}

Analyze this vulnerability and explain the risk.

// Step 3: Our code sends prompt to LLM

const analysis = await callGemini(prompt)

The Backend is the Middleman:

Fetches real data from APIs (NVD, CISA, GitHub)

Packages that data into a prompt

Sends prompt to LLM

Returns LLM's response to the user

🤖

Multi-Agent: What It Really Means

⚡Honest Assessment

In SecPhi's current form, "multi-agent" is primarily a presentation layer. Here's what's actually happening versus what it looks like.

What It Looks Like	What It Actually Is
4 specialized agents debating	Same LLM called 4 times with different prompts
Agents reaching consensus	Your code averaging their outputs
Real-time debate	Sequential API calls displayed with animation
Different expert perspectives	Different system prompts to same model

What the Code Actually Does:

// "Multi-agent" = same LLM with different prompts

// "Scanner Agent"

const scanner = await callGemini("You are a scanner. Analyze...")

// "Scorer Agent"

const scorer = await callGemini("You are a risk scorer. Rate...")

// "Exploit Agent"

const exploit = await callGemini("You are a threat expert. Assess...")

// "Patch Agent"

const patch = await callGemini("You are a remediation expert. Fix...")

💬

A single well-written prompt could produce 80% of the same value. The multi-agent UI adds presentation clarity and demonstrates architectural thinking.

🎯

What "Real" Multi-Agent Would Look Like

📍SecPhi Today

All agents use same NVD data
Agents run in parallel, never see each other
Same Gemini model with different prompts
Consensus = averaging scores

🚀Real Multi-Agent

Each agent queries DIFFERENT data sources
Agents respond to each other's outputs
Different specialized models
Agents can disagree and flag conflicts

✅

Why We Built It This Way

🏗️

Architecture Ready

The structure is in place to evolve into true multi-agent

📊

Verified Data

CVE info comes from official sources, not AI hallucinations

🎨

Clear UX

Tabbed agent interface makes complex analysis digestible

💡

Educational

Demonstrates multi-agent concepts for learning & interviews

Transparency Builds Trust

We believe in being honest about what AI can and can't do. That's how we build products you can actually rely on.

Explore Full Architecture Try SecPhi