System Performance

Live session metrics · architecture overview · agent roster

speed Session Metrics

—

Queries

—

Avg Latency (ms)

—

Debate Queries

—

Uptime (min)

Query Mode Distribution

Recent Query Latencies (ms)

System Architecture

User Query

↓

Local Retrieval

ChromaDB · all-MiniLM-L6

10,681 chunks · 19 papers

↔

Web Fallback

DuckDuckGo (primary)

Tavily (fallback)

sim < 0.45 OR recency keywords → web triggered

↓

Multi-Agent Swarm Debate

Alpha

Llama 3.3 70B

Skeptic

Beta

GPT-OSS 120B

Literalist

Gamma

Llama 4 Scout

Connector

Delta

Qwen3-32B

Reasoner

Round 1: independent answers · Round 2+: argue · convergence check

↓

Epsilon — Synthesizer

GPT-OSS 20B · Groq · reconciles debate → final answer

↓

Citation Enrichment

Semantic Scholar API arXiv API OpenAlex API PubMed E-utils

↓

Final Answer + Sources + Debate Transcript

Agent Roster

Alpha - Skeptic groq

llama-3.3-70b-versatile

Challenges every claim aggressively. Demands evidence. Points out what is NOT in the sources.

Beta - Deep Reasoner groq

openai/gpt-oss-120b

Thinks step by step. Shows chain of thought. Identifies hidden assumptions and logical gaps.

Gamma - Connector groq

meta-llama/llama-4-scout-17b-16e-instruct

Finds non-obvious connections between sources. Thinks laterally across papers.

Delta - Literalist groq

qwen/qwen3-32b

Accepts only what is explicitly stated. Rejects all inferences and extrapolations.

Epsilon - Synthesizer groq

openai/gpt-oss-20b · Fast synthesis of full debate transcript

Reconciles all agent arguments. Identifies consensus. Produces final authoritative answer with citations and confidence rating.

database Corpus

—

Chunks

—

Papers

0.45

Fallback Threshold

—

Corpus Status

Indexed Papers