Perspectra

Think clearly. See the whole picture.
A local-first, privacy-preserving analyzer for social media claims, ragebait, and echo chambers across English, Hindi, and Hinglish. It exposes underlying rhetorical tactics and constructs a steelmanned opposing perspective with the evidence required to settle the dispute.

Detail	Specification
Track	Narrow Agentic Tools (Local Inference Agents)
Event	Hacktoberfest Hack Day Jaipur × Hack IT Sapiens (LNMIIT)
Repository	github.com/Ritwik82/Perspectra
Licence	Apache-2.0

---

1. Problem Statement & Target User

The Problem

Social media feeds (X/Twitter, Reddit, Instagram, LinkedIn) are optimized for engagement through one-sided framing, outrage cultivation, and selective omission. Even when a viral post is factually sound or widely applauded, it is frequently one-sided—swapping complex tradeoffs for tribal certainty. Readers are trapped inside echo chambers not because they are uncritical, but because alternative perspectives are omitted or presented as weak strawmen.

Target User

• Everyday readers & students seeking to evaluate controversial claims or viral screenshots without being manipulated by emotional framing.
• Researchers & critical thinkers who need an objective, two-sided breakdown of complex debates in multilingual contexts (English, Hindi, and code-switched Hinglish).
• Privacy-conscious users who refuse to upload their private screenshots, reading history, or personal browsing habits to proprietary cloud APIs.

---

2. Approach, Architecture & Workflow

Perspectra is designed as a Narrow Agentic Tool: it does one job end-to-end with high precision. It accepts either raw text or screenshot images, extracts text via multi-modal vision OCR, analyzes rhetorical tactics, scores reasoning quality, and constructs a steelmanned counter-argument.

Core Architecture

┌─────────────────────────────────────────────────────────────┐
│                 Browser Client (Next.js 16)                 │
│   • Textarea paste / drag-and-drop screenshot picker        │
│   • Multi-mode selector (Quick, Deep, Counter)              │
│   • Dynamic source quotation & citation highlighting        │
│   • Live elapsed execution timer & responsive dual-pane     │
└──────────────────────────────┬──────────────────────────────┘
                               │ POST /api/analyze
                               ▼
┌─────────────────────────────────────────────────────────────┐
│            Next.js Server-Side Route Handler                │
│   • Input validation, sanitization & payload limits         │
│   • Zero browser-to-Ollama CORS leaks                       │
│   • Timeout controller (300s) & context retry watchdog      │
└──────────────────────────────┬──────────────────────────────┘
                               │ POST /api/chat
                               ▼
┌─────────────────────────────────────────────────────────────┐
│                Local Ollama Inference Engine                │
│   • Model: gemma4:e2b (4.6B, Q4_K_M)                        │
│   • num_ctx: 16384 (ensures complete generation)            │
│   • temperature: 0.2 (deterministic calibration)            │
│   • Strict JSON Schema enforcement (analysisSchema)         │
└──────────────────────────────┬──────────────────────────────┘
                               │ Structured JSON Output
                               ▼
┌─────────────────────────────────────────────────────────────┐
│                 ReportCard & OtherSidePanel                 │
│   • What it claims (distilled thesis)                       │
│   • What it's doing (identified tactics + exact quotes)     │
│   • THE OTHER SIDE (steelmanned counter-view)               │
│   • WHAT WOULD SETTLE IT (empirical deciding test)          │
│   • Reasoning Score & Categorical Verdict Badge             │
│   • Confidence metric & rationale                           │
└─────────────────────────────────────────────────────────────┘

Guardrails & Failure Handling

• The "No Manipulation" Guardrail: Many argument analyzers force-fit fallacies onto innocent text. Perspectra actively monitors for exploratory, neutral, or self-hedging text. In these cases, it marks noSignal: true, emits an empty tactics list, and refrains from fabricating attacks. Even with neutral posts, the counter-perspective is provided to ensure intellectual breadth.
• Truncation Watchdog: Large system prompts swallow default context buffers. Perspectra enforces num_ctx = 16384 and inspects done_reason. If a model response fails to terminate with "stop", the runtime triggers an automatic fallback to 32768.
• Zero Cloud Guarantee: All operations execute completely offline on local hardware with Wi-Fi disabled.

---

3. Model Details

Attribute	Value
Model Name	Google Gemma 4 E2B (gemma4:e2b)
Parameter Count	4.6 Billion (4.6B)
Quantization	Q4_K_M (4.58 GB on disk)
Model Licence	Gemma Terms of Use (Open Weights)
Local Runtime	Ollama 0.40+ (llamacpp runner)
Context Window	16,384 tokens (num_ctx: 16384)
Capabilities Used	Completion, Multimodal Vision (OCR + Screenshot reading)
External APIs	None. Zero hosted frontier APIs (no OpenAI, Anthropic, or Gemini API keys).
Target Hardware Tested	NVIDIA GeForce RTX 2050 (4 GB VRAM), 16 GB DDR4 RAM, Windows 11

---

4. Setup & Running Locally

1. Prerequisites

• Node.js v18.18+ or v20+
• Ollama installed and running (ollama.com)
• Python 3.10+ (optional, for running benchmark verification scripts)

2. Pull the Local Model

ollama run gemma4:e2b

(Verify it responds, then you can exit the chat with /bye)

3. Clone & Install Dependencies

git clone https://github.com/Ritwik82/Perspectra.git
cd Perspectra
npm install

4. Configure Environment

Copy the example environment configuration:

cp .env.example .env.local

(Defaults point to http://localhost:11434 with model gemma4:e2b and OLLAMA_NUM_CTX=16384)

5. Launch the Application

npm run dev

Open http://localhost:3000 in your browser.

6. Example Test Inputs

• English Text: Paste an excerpt arguing about workplace policies or the gender pay gap (tests/01_text_en.txt).
• Hinglish Screenshot: Drop an Instagram or Reddit screenshot discussing everyday lifestyle, diets, or fitness (tests/04_img_en.jpeg).
• Hindi Devanagari: Test a civic policy or transport debate in pure Devanagari script (tests/06_neutral_hi.txt).

---

5. Benchmark Results & Evidence

Tested across multiple runs on target laptop hardware (NVIDIA RTX 2050, 4 GB VRAM) using tests/run_all.py:

Test Input	Input Type	Language	Output Tokens	Latency	Speed	Verdict / Assessment	Done Reason
01_text_en.txt	Text	English	892 tok	87.3 s	23.6 tok/s	Nuanced critique, calls for transparency	stop
03_text_hinglish.txt	Text	Hinglish	814 tok	71.2 s	37.8 tok/s	Personal observation, hedged claims	stop
05_img_hi.jpeg	Screenshot	Hindi (Devanagari)	948 tok	65.5 s	39.8 tok/s	Political claim, rhetorical questions	stop
06_neutral_eng.txt	Text	English	722 tok	77.6 s	39.9 tok/s	No Signal (Guardrail passed)	stop
08_img_en.jpeg	Screenshot	English	856 tok	67.1 s	42.4 tok/s	Anecdotal narrative, weak evidence	stop

Known Limitations

1. Model Capacity (4.6B): While Gemma 4 E2B shows high reasoning quality, subtle sarcasm or deeply localized idioms occasionally require the deeper prompt mode.
2. Real-Time Fact-Checking: Perspectra evaluates internal argument consistency, evidential structure, and rhetorical framing. It does not browse the live internet to verify breaking news facts.
3. Hardware Latency: On low-power 4 GB VRAM GPUs, vision-based full OCR inference takes 60–80 seconds. Warmed-up runs typically complete faster.