Skip to content

PROMPTS.md

Latest commit

 

History

History
247 lines (169 loc) · 9.6 KB

PROMPTS.md

File metadata and controls

247 lines (169 loc) · 9.6 KB

Pinecone Assistant MCP - Prompt Templates

This document details the prompt templates included with the Pinecone Assistant MCP Server for guided document research workflows.

⚡ Prompt Templates (Corpus-Neutral Research Workflows)

The Pinecone Assistant MCP includes 4 prompt templates that appear in the Claude Desktop UI. These templates provide structured research workflows optimized for token efficiency and eliminate the need to manually construct tool calls.

🎯 How to Use Prompt Templates

  1. In Claude Desktop: Press the + button

    Step 1 - Select prompt template

  2. Select "Add from pinecone_assistant" from the dropdown menu

    Step 2 - Select Pinecone Assistant MCP

  3. Select the desired Prompt from the dropdown menu

    Step 3 - Select desired prompt

  4. Fill in the parameters — read the guidance for which fields are required

    Step 4 - Fill in parameters

  5. Execute - (Recommendation: Type in "Perform the Prompt" before hitting Execute) — Claude will automatically run the complete workflow

    Step 5 - Execute workflow

Overview

All four prompts are corpus-neutral — they work with any document domain (legal, medical, technical, financial). They are pre-configured to use context-only tools by default, minimizing token costs on free-tier plans.

Token Tiers

Tier Tools Used Token Pool Cost Impact
Context only assistant_context, assistant_strategic_multi_search_context Context tokens Lowest cost
Context + LLM assistant_chat Context + input/output tokens Higher cost

Free Tier Note: Context-only prompts draw from the 500K context token lifetime pool. LLM prompts draw from the 1.5M input / 200K output lifetime pools. See README for full pricing details.

Research Prompts

deep_research — Thorough Multi-Angle Coverage

Purpose: Comprehensive research across multiple search angles using strategic multi-search patterns.

Token Tier: Context only (no AI synthesis cost)

Parameters:

Parameter Required Description
topic Required The research topic or question
domain Optional Strategic search domain from strategic-searches.yaml (e.g., section_101_eligibility for USPTO corpus)

Generated Workflow:

  1. Calls get_configuration_status to confirm active assistant and available domains
  2. Executes assistant_strategic_multi_search_context with the topic and optional domain
  3. Returns raw document chunks organized by search pattern for Claude to analyze
  4. Claude synthesizes findings from retrieved content

Best For:

  • Comprehensive topic coverage when you need multiple angles
  • Research before drafting responses or analysis documents
  • Discovery phase — understanding what documents say before deeper dives

Example Use:

Topic: "obviousness combination of references"
Domain: section_103_obviousness

quick_lookup — Fast Single-Fact Retrieval

Purpose: Rapid lookup of a specific fact, rule, or concept with a single tool call.

Token Tier: Context only (no AI synthesis cost)

Parameters:

Parameter Required Description
topic Required The specific topic, rule number, or concept to look up

Generated Workflow:

  1. Calls assistant_context with top_k=3, snippet_size=1024 for minimal token usage
  2. Returns the most relevant document chunks matching the query
  3. Claude presents findings from retrieved content

Best For:

  • Looking up a specific rule, statute, or procedure
  • Verifying a single fact quickly
  • When you know exactly what you need and don't need broad coverage
  • Conserving context token budget on the free tier

Example Use:

Topic: "MPEP 2106 Alice framework step 2A prong 1"

comparative_research — Side-by-Side Comparison

Purpose: Retrieve and compare documentation on two related topics simultaneously.

Token Tier: Context only (no AI synthesis cost)

Parameters:

Parameter Required Description
topic_a Required First topic to research and compare
topic_b Required Second topic to research and compare

Generated Workflow:

  1. Calls assistant_context for topic_a (top_k=5)
  2. Calls assistant_context for topic_b (top_k=5)
  3. Returns both sets of document chunks
  4. Claude analyzes and presents a structured comparison

Best For:

  • Comparing two legal standards, procedures, or concepts
  • Understanding differences between related rules or exceptions
  • Due diligence requiring side-by-side analysis
  • Comparing how similar concepts are treated in different contexts

Example Use:

Topic A: "IPR estoppel under 35 USC 315(e)"
Topic B: "PGR estoppel under 35 USC 325(e)"

delegated_research — AI-Synthesized Answers via Delegation

Purpose: Delegate research and synthesis to the Pinecone Assistant's AI, receiving a compact citation-backed answer. Pinecone handles retrieval internally — Claude receives only the synthesized result, preserving Claude's context window.

Token Tier: Context + LLM (uses input/output token pools — see free tier warnings)

Parameters:

Parameter Required Description
research_question Required The full research question for the Pinecone AI to answer
model Optional AI model to use (default: gpt-4o). Options: gpt-4o, gpt-4.1, o4-mini, claude-3-5-sonnet, claude-3-7-sonnet, gemini-2.5-pro
prior_context Optional Compact summary of prior conversation context to include for continuity (avoid full conversation history)

Generated Workflow:

  1. Calls assistant_chat with the research question
  2. Pinecone internally retrieves relevant document chunks and feeds them to the configured LLM
  3. Returns a synthesized, citation-backed answer (~500–2000 tokens)
  4. Claude presents the answer without re-summarizing

Best For:

  • Questions requiring AI synthesis across multiple document sections
  • When Claude's context window is under pressure (delegation returns compact results)
  • Agentic workflows chaining multiple independent research questions
  • Paid plan users where token cost is less of a concern

⚠️ Free Tier Warning: Each call consumes 30K+ input tokens plus output tokens from lifetime pools. Use deep_research or quick_lookup for context-only retrieval when possible.

Example Use:

Research question: "What are the requirements for establishing a nexus between objective indicia of nonobviousness and the claimed invention under MPEP 716?"
Model: gpt-4o
Prior context: (leave empty for independent questions)

Agentic Chaining Pattern (for multi-part research):

# Each delegated_research call is stateless and independent
Call 1: "What does MPEP say about claim construction for MPF claims?"
Call 2: "What are the indefiniteness standards under 35 USC 112(b)?"
Call 3: "How does the examiner apply the broadest reasonable interpretation?"

# Claude synthesizes all three compact answers

Prompt Selection Guide

I want to... Use This Prompt Token Cost
Research a topic thoroughly from multiple angles deep_research Low
Look up a specific rule or fact quickly quick_lookup Lowest
Compare two related concepts side by side comparative_research Low
Get a synthesized AI answer with citations delegated_research High

Decision Flow

Do I need AI synthesis?
├── No → Use context-only prompts (deep_research, quick_lookup, comparative_research)
│   ├── Single specific fact → quick_lookup
│   ├── Two topics to compare → comparative_research
│   └── Broad coverage needed → deep_research
│
└── Yes → Use delegated_research
    ├── On free tier? → Consider context-only prompts instead
    └── On paid plan or agentic workflow? → delegated_research is ideal

Integration with Claude Skills

The prompt templates work best when combined with the included Claude skills, which provide additional guidance on tool selection and workflow optimization:

Skill Best Prompts Notes
pinecone-assistant All prompts Generic corpus research guidance
pinecone-assistant-uspto deep_research with USPTO domains Patent law domain selection and workflows
pinecone-assistant-paid-plan delegated_research Delegation patterns and context_options tuning

See skills/README.md for skill installation and usage.

Cross-MCP Integration

When used with the other USPTO MCP servers, prompts can be chained to build comprehensive research workflows:

Example: Office Action Response Research

  1. Use quick_lookup to find the relevant MPEP section for the rejection type
  2. Use deep_research with the appropriate domain for comprehensive guidance
  3. Switch to uspto_pfw_mcp to pull the actual prosecution history
  4. Use delegated_research to synthesize guidance into a draft response framework

For detailed integration patterns, see USAGE_EXAMPLES.md.

For setup and installation guidance, see README.md and INSTALL.md.

For API key setup, see PINECONE_KEY_GUIDE.md.