How QphiQ 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:
LLMs don't browse - they only see what's in the prompt
They can't fetch data from databases or external services
Files must be converted to text and put in the prompt
Your application code must handle data retrieval
LLMs only "see" what your application code feeds them. When QphiQ shows AI analysis, the AI isn't fetching data—our backend code does, then packages it into a prompt for the AI to analyze.
What Actually Happens
// Step 1: Our code makes API calls
const openAlexData = await fetch('https://api.openalex.org/works?search=CRISPR')
const semanticData = await fetch('https://api.semanticscholar.org/...')
// Step 2: Our code handles cross-validation WITH real data
const crossValidated = mergePapers(openAlexData, semanticData)
// Step 3: Our code sends data to LLM for synthesis
const synthesis = await anthropic.messages.create({
model: 'claude-3-haiku',
messages: [{ role: 'user', content: `Analyze these papers: ${crossValidated}` }]
})
// Step 4: Return combined results
return { papers: crossValidated, analysis: synthesis }The Backend is the Middleman:
Fetches raw data from APIs (OpenAlex, Semantic Scholar, SEC EDGAR)
Packages that data into a prompt
Sends prompt to LLM
Returns LLM's response to the user
Multi-Agent: What It Really Means
In QphiQ's current form, "multi-agent" is primarily a presentation layer. Here's what's actually happening versus what it looks like:
- • 4 specialized agents debating
- • Agents reaching consensus
- • Real-time debate
- • Different expert perspectives
- • Same LLM called 4 times with different prompts
- • Your code averaging their outputs
- • Sequential API calls displayed with animation
- • Different system prompts to same model
What the Code Actually Does:
// "Multi-agent" = same LLM with different prompts
const literatureAgent = await claude("You are a literature expert. Analyze...")
const trendsAgent = await claude("You are a trends analyst. Analyze...")
const authorAgent = await claude("You are an author expert. Analyze...")
const verifyAgent = await claude("You are a verification expert. Check...")
A single well-written prompt could produce 80% of the same value. The multi-agent UI adds presentation clarity and demonstrates architectural thinking—valuable for showing engineering sophistication to enterprise buyers.
- •All agents use same RAW data
- •Agents can't update state or react to each other
- •Same base LLM model with different prompts
- •"Consensus" = averaging scores
- •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 when compute costs allow. Current implementation demonstrates the pattern.
Verified Data
100% of claims come from official sources, not AI hallucination. Cross-validation happens at the data layer, not the LLM layer.
Clear UX
Nested agent interface enables complex analysis. Agent-like abstractions give users mental models to navigate.
Educational
Shows enterprise buyers multi-agent concepts in action. Demonstrates deep understanding of AI architecture.
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.