prism-intent 0.1.1

PRISM: Predicting Response Intent from Session Memory

PRISM: Predicting Response Intent from Session Memory

PRISM is a model-agnostic Python library and research framework designed to eliminate the "Intent Alignment Tax" in AI conversations. By analyzing past chat history, PRISM builds a User Intent Vector (UIV) that predicts a user's preferred style, format, and complexity, allowing LLMs to provide the perfect response on the first try.

🚀 Key Features

• Clarification Detection: Automatically identify when users are frustrated by misaligned AI responses.
• Dynamic User Profiling: Build long-term memory of user preferences (e.g., "likes bullet points", "prefers simple analogies").
• Seamless Integration: Optimized for Llama 3.2:3b via Ollama, but works with OpenAI, Anthropic, and other LLMs.
• Interactive Web UI: Real-time visualization of intent extraction and prompt adaptation using Gradio.
• Research Framework: Tools to evaluate intent alignment on the OpenAssistant/oasst1 dataset.

📦 Installation

# Install core dependencies
pip install datasets pandas ollama gradio

🧩 CLI Plugin (OpenAI/Claude/Gemini)

The CLI generates provider-ready payloads with PRISM injection so you can plug it into any API client.

# Install the package locally to get the `prism` command
pip install -e .

# Create an OpenAI-ready payload from a single user message
prism adapt --provider openai --message "Explain vector databases"

# Use a history file and output only the provider payload JSON
prism adapt --provider anthropic --history examples/history.json --message "Summarize in bullets" --payload-only

Notes

• Defaults: OpenAI gpt-4o-mini, Anthropic claude-3-5-sonnet-latest, Gemini gemini-1.5-flash (override with --model).
• Disable injection globally with PRISM_DISABLE_PERSONALIZATION=1.
• Output JSON includes the UIV and confidence unless you pass --payload-only.

CLI Demo (Before/After)

Before	After

🔌 Provider Integration

PRISM outputs provider-ready payloads. Pipe them into your own API client.

# Create a payload JSON for any provider
prism adapt --provider gemini --history examples/history.json --message "Summarize in bullets" --payload-only > payload.json

🏷️ Recommended GitHub Topics

llm · personalization · prompt-engineering · nlp · rag · chatbots · gradio · ollama · gemini · claude · openai · ai-tools

🎥 Demo in Action: Visual Tour

PRISM eliminates the "Intent Alignment Tax" in 6 clear steps:

Step 1: The Problem	Step 2: The Correction	Step 3: PRISM Analysis
User receives a generic, long response	User provides a stylistic correction	PRISM extracts the Intent Vector (UIV)

Step 4: Persistence	Step 5: The Solution	Step 6: Local Integration
Preferences saved to JSON Profile Store	New queries get the right style immediately	Fully powered by Local Llama 3.2:3b

---

The "Intent Alignment" Workflow:

1. User: "Explain Quantum Computing" -> AI: [Long Technical Paragraph]
2. User: "Too complex, use bullets" -> PRISM: Detects intent and updates UIV to {'complexity': 'simple', 'format': 'bullets'}.
3. User: "What is a Qubit?" -> AI (via PRISM): [Automatically provides a simple, bulleted explanation on the first try].

---

🏗️ Architecture

PRISM sits between the User and the LLM (Ollama/Llama 3.2), acting as a persistent memory and intent-alignment layer.

graph LR
    A[User] --> B[PRISM Intent Detector]
    B --> C[UIV Builder]
    C --> D[Local Profile Store]
    D --> E[Ollama Llama 3.2]
    E --> A

📄 Research Paper

This project implements the methodology described in the PRISM Research Paper.

✅ Production Readiness v1 (Implemented)

PRISM now includes a full production-hardening pass across six engineering phases:

1. Hybrid Intent Inference Core
   • Weighted hybrid signals (regex + multi-turn preference scoring)
   • Confidence estimation + abstain gating
   • Temporal decay blending with prior profile
2. Profile Lifecycle Controls
   • Profile schema v2 (uiv, confidence metadata, source, override)
   • View/edit/reset support in storage APIs
   • Temporary override with TTL
3. Observability
   • Structured runtime event logging (data/metrics/events.jsonl)
   • Decision counters (data/metrics/counters.json)
   • Wrong-personalization signal tracking
4. Evaluation Upgrade
   • Production metrics tracker (clarification_rate, first_response_acceptance_rate, wrong_personalization_rate, etc.)
   • Extended research pipeline output to research\outputs\evaluation_metrics.json
5. Storage Hardening
   • Profile versioning and metadata normalization
   • Export/delete profile methods for privacy workflows
   • Retention policy hook (apply_retention_policy)
   • Store abstraction interface for DB migration path
6. Release Guardrails
   • Rollback switch: set PRISM_DISABLE_PERSONALIZATION=1 to disable injection globally
   • Drift checks via runtime counters (correction/abstain rate alerts)

Updated Runtime Architecture

graph LR
    A[User] --> B[Hybrid Intent Inference]
    B --> C[Confidence + Abstain Gate]
    C --> D[Profile Store v2]
    D --> E[Prompt Adapter]
    E --> F[LLM]
    F --> A
    E --> G[Telemetry + Drift Monitor]

Running Research Evaluation

python research/01_evaluate_lmsys.py --num-samples 10000 --dataset-name OpenAssistant/oasst1

📄 License

MIT