Skip to content

USAGE_EXAMPLES.md

Latest commit

 

History

History
789 lines (602 loc) · 27 KB

USAGE_EXAMPLES.md

File metadata and controls

789 lines (602 loc) · 27 KB

Pinecone Assistant MCP - Usage Examples

This document provides comprehensive examples and guidance for using the Pinecone Assistant MCP effectively, with token optimization strategies and best practices.

Note: This guide uses USPTO patent examination as the reference implementation, but the patterns apply to any document corpus.

Multi-turn Conversation Token Management

CRITICAL: The Pinecone Assistant API is stateless - it has no memory between requests.

When to Include Conversation History ✅

Only include history when the new question explicitly references prior context:

  • Pronouns: "that", "this", "it", "those", "these"
  • References: "as mentioned", "the above", "previously discussed"
  • Continuation: "follow up", "expand on", "additionally"

Example requiring history:

{
  "tool": "assistant_chat",
  "arguments": {
    "messages": [
      {"role": "user", "content": "What is the Alice/Mayo framework?"},
      {"role": "assistant", "content": "[Previous AI response]"},
      {"role": "user", "content": "How does THAT framework apply to ML patents?"}
    ]
  }
}

When NOT to Include History ❌

Do NOT include history when questions are independent:

Bad (wasteful):

{
  "messages": [
    {"role": "user", "content": "What is Section 101?"},
    {"role": "assistant", "content": "[1000 token response]"},
    {"role": "user", "content": "What is Section 103?"}  // Independent question!
  ]
}

Good (efficient):

{
  "messages": [
    {"role": "user", "content": "What is Section 103?"}  // Stateless
  ]
}

Token Impact Example

4-turn conversation comparison:

  • Wasteful (full history): 130K input tokens total
  • Smart (stateless when possible): 124K input tokens total
  • Savings: 6K tokens = 3-5 additional chat queries from your lifetime allocation

Default pattern: Send single-message queries. Only add history if the new question literally cannot be understood without prior context.

Temperature Configuration

Default: 0.2 (Low Temperature for Legal/USPTO Precision)

The MCP is configured with a low default temperature (0.2) optimized for patent law and legal analysis:

Why Low Temperature for USPTO/Legal Work:

  • Accurate Citations: Prevents AI from inventing MPEP sections or case names
  • Consistent Analysis: Same query yields consistent legal interpretations
  • Factual Precision: Critical for exact statute/regulation references
  • Risk Mitigation: Reduces hallucinations in legal advice
  • IRAC Reliability: Structured analysis requires deterministic responses

Temperature Guide:

Temperature Behavior Best For USPTO Suitability
0.0 - 0.3 Deterministic, precise Legal citations, factual analysis ✅ Recommended
0.4 - 0.7 Balanced General research ⚠️ Acceptable for exploratory research
0.8 - 2.0 Creative, varied Brainstorming, creative writing ❌ Not recommended for legal work

Override Temperature:

{
  "tool": "assistant_chat",
  "arguments": {
    "messages": [...],
    "temperature": 0.3  // Override default for this specific query
  }
}

Custom Domains: Set higher temperature for creative domains via environment variable:

export DEFAULT_TEMPERATURE=0.7  # For creative/brainstorming domains

Free Tier Budget Management

Lifetime Token Allocation:

  • Input tokens: 1.5M
  • Context tokens: 500K
  • Output tokens: 200K

Usage estimates (total lifetime, not per month — these do NOT reset):

  • assistant_context: ~50+ total queries (context tokens)
  • assistant_strategic_multi_search_context: ~20-30 total queries (context tokens)
  • assistant_strategic_multi_search_chat: ~50-75 total queries (input tokens)
  • assistant_chat (with smart history): ~50 total queries (input tokens)
  • assistant_chat (with wasteful history): ~35 total queries (input tokens)

🎯 Domain Organization & Use Cases

Domain Selection Guide

Domain Focus Area Typical Use Cases Search Patterns
section_101_eligibility Patent eligibility analysis Alice/Mayo rejections, abstract idea analysis 4 searches
section_103_obviousness Obviousness analysis KSR factors, prior art combination 4 searches
section_112_requirements Specification compliance Written description, enablement, definiteness 4 searches
section_102_novelty Prior art and anticipation Novelty rejections, prior art dates 3 searches
claim_construction Claim interpretation Phillips standard, prosecution history 3 searches
ptab_procedures Post-grant proceedings IPR/PGR, PTAB claim construction 3 searches
mechanical_patents Mechanical inventions Mechanical obviousness, manufacturing 2 searches
software_patents Software/AI inventions Software eligibility, AI/ML patents 2 searches
general_patent_law General guidance Broad patent law questions 5 searches

🔧 Domain Design Rationale

Domains are organized by legal issue rather than workflow stage, enabling:

  • Direct issue targeting: Select domain matching your specific legal question
  • Cross-reference capability: Combine multiple domains for complex analysis
  • Technology-specific searches: software_patents and mechanical_patents for tech-focused guidance

Selection Strategy:

  • Specific rejection type → Use matching section domain (101, 102, 103, 112)
  • Technology-specific → Use software_patents or mechanical_patents
  • General research → Start with general_patent_law
  • Post-grant → Use ptab_procedures or claim_construction

📚 Detailed Usage Scenarios

Scenario 1: Section 101 Rejection (Software Patent)

Situation: Received Alice/Mayo rejection on machine learning algorithm patent

{
  "tool": "assistant_strategic_multi_search_chat",
  "arguments": {
    "query": "machine learning algorithm patent eligibility",
    "domain": "section_101_eligibility",
    "max_searches": 3
  }
}

Searches Executed:

  • Alice/Mayo framework analysis
  • Technological improvement vs abstract idea
  • Inventive concept analysis

Workflow:

  1. Strategic search → Get Alice/Mayo framework and practical application guidance
  2. PFW MCP → Pull office action with 101 rejection
  3. Draft response using MPEP examples and case law

Expected Results:

  • Alice/Mayo two-step framework application
  • Technological improvement arguments
  • Inventive concept analysis for Step 2

Scenario 2: Office Action Response (103 Rejection)

Situation: Need to overcome obviousness rejection for pharmaceutical formulation

{
  "tool": "assistant_strategic_multi_search_chat",
  "arguments": {
    "query": "pharmaceutical formulation prior art combination",
    "domain": "section_103_obviousness",
    "max_searches": 4
  }
}

Searches Executed:

  • Graham factors analysis
  • KSR motivation to combine rationales
  • Secondary considerations of non-obviousness
  • Prior art combination analysis

Workflow:

  1. PFW MCP → Get office action and prior art references
  2. Strategic search → Find KSR rationales and secondary considerations guidance
  3. Draft response with Graham factors and secondary considerations arguments

Expected Results:

  • Graham v. John Deere four-factor framework
  • KSR motivation to combine analysis
  • Secondary considerations arguments (commercial success, teaching away)

Scenario 3: IPR Institution Decision Analysis

Situation: Analyzing likelihood of IPR institution for challenged patent

{
  "tool": "assistant_strategic_multi_search_chat",
  "arguments": {
    "query": "IPR institution standards claim construction",
    "domain": "ptab_procedures",
    "max_searches": 3
  }
}

Searches Executed:

  • IPR petition and institution standards
  • PTAB claim construction standards
  • PTAB estoppel rules

Workflow:

  1. Strategic search → Get institution standards and claim construction guidance
  2. PTAB MCP → Search similar IPR proceedings and institution decisions
  3. Analyze likelihood of institution based on precedent

Expected Results:

  • IPR institution standards and preponderance of evidence requirements
  • PTAB claim construction methodology (BRI/Phillips)
  • Estoppel implications for non-instituted grounds

Scenario 4: Written Description Rejection

Situation: Responding to Section 112(a) written description rejection

{
  "tool": "assistant_strategic_multi_search_chat",
  "arguments": {
    "query": "written description possession requirement biotechnology",
    "domain": "section_112_requirements",
    "max_searches": 2
  }
}

Searches Executed:

  • Written description requirement analysis
  • Enablement requirement and Wands factors

Workflow:

  1. Strategic search → Get written description and enablement standards
  2. PFW MCP → Pull specification and office action
  3. Draft response with possession and enablement arguments

Expected Results:

  • Written description possession requirement (§112(a))
  • Enablement standards and Wands factors
  • Specification support analysis

Scenario 5: Claim Construction Dispute

Situation: Analyzing claim scope for litigation or licensing

{
  "tool": "assistant_strategic_multi_search_chat",
  "arguments": {
    "query": "claim construction means-plus-function limitations",
    "domain": "claim_construction",
    "max_searches": 3
  }
}

Searches Executed:

  • Phillips v. AWH claim construction standard
  • Prosecution history estoppel analysis
  • Functional claim language interpretation

Workflow:

  1. Strategic search → Get claim construction standards and Phillips analysis
  2. PFW MCP → Review prosecution history for estoppel issues
  3. Analyze claim scope using intrinsic and extrinsic evidence

Expected Results:

  • Phillips standard for claim construction
  • Prosecution history estoppel principles
  • Means-plus-function interpretation under §112(f)

📝 MCP Prompts

Four corpus-neutral prompt templates are accessible from the Claude prompt menu. Each generates a pre-filled workflow with structured tool calls — select from the prompt menu and fill in the arguments.

Prompt Parameters Token Tier Use When
deep_research topic, domain? Context only Thorough multi-angle coverage
quick_lookup topic Context only Fast single-fact retrieval
comparative_research topic_a, topic_b Context only Side-by-side topic comparison
delegated_research research_question, model?, prior_context? Context + LLM Paid plan / agentic synthesis

deep_research

Executes assistant_strategic_multi_search_context with max_searches=4, then fills gaps with targeted assistant_context calls. Context tokens only — no AI synthesis cost.

Example: topic="Alice/Mayo two-step framework software patents", domain="section_101_eligibility"

quick_lookup

Single assistant_context call with top_k=3, snippet_size=1024. Under 5K context tokens. One retry with broader terminology if initial results are irrelevant.

Example: topic="MPEP 2106.05(a) practical application"

comparative_research

Two sequential assistant_context calls — one per topic — then structured comparison output covering similarities, differences, and when to apply each.

Example: topic_a="written description requirement", topic_b="enablement requirement"

delegated_research

Delegates synthesis to the Pinecone Assistant AI via assistant_chat. Pinecone handles internal retrieval and synthesis; Claude receives only the compact result. Supports stateless single queries and follow-up queries with prior context.

Example: research_question="What are the key differences between IPR and PGR institution standards?", model="gpt-4o"

Follow-up with prior context: Paste the prior assistant answer into prior_context when the new question references it. Omit prior_context for independent questions (stateless = cheapest).

🔀 Delegation Workflow (Paid Plan / Agentic)

assistant_chat functions as a sub-agent delegation mechanism — Pinecone internally retrieves context from the knowledge base, feeds it to the configured LLM, and returns a synthesized, citation-backed answer. Claude receives only the compact result (~500–2000 tokens) rather than raw document chunks, preserving Claude's context window for orchestration.

Single Delegated Query

{
  "tool": "assistant_chat",
  "arguments": {
    "messages": [{"role": "user", "content": "What are the seven KSR rationales for combining prior art references?"}],
    "model": "gpt-4o",
    "temperature": 0.2,
    "include_highlights": true,
    "context_options": {"top_k": 5, "snippet_size": 2048}
  }
}

Agentic Chaining (Multiple Independent Questions)

Chain independent calls without history — each is stateless, no accumulated cost:

// Call 1: First research topic
{"tool": "assistant_chat", "arguments": {"messages": [{"role": "user", "content": "IPR institution standards"}]}}

// Call 2: Second independent topic — no history
{"tool": "assistant_chat", "arguments": {"messages": [{"role": "user", "content": "PGR institution standards"}]}}

// Synthesize both results in Claude's context

Follow-Up With Prior Context

Include history only when the new question explicitly references the prior answer:

{
  "tool": "assistant_chat",
  "arguments": {
    "messages": [
      {"role": "user", "content": "What are the Graham v. John Deere four factors?"},
      {"role": "assistant", "content": "[prior answer text]"},
      {"role": "user", "content": "How do those factors apply to pharmaceutical formulations?"}
    ],
    "context_options": {"top_k": 5, "snippet_size": 2048}
  }
}

context_options Tuning

Control how much context Pinecone sends to its LLM per call:

Setting Tokens to LLM Use When
{"top_k": 3, "snippet_size": 1024} ~3K Focused single-concept question
{"top_k": 5, "snippet_size": 2048} ~10K Standard research (recommended)
{"top_k": 10, "snippet_size": 2048} ~20K Broad topic needing wide coverage

Mixing Delegation with Context Retrieval

Use assistant_chat for broad synthesis, then assistant_context for targeted follow-up gaps:

// Step 1: Broad synthesis via delegation
{"tool": "assistant_chat", "arguments": {"messages": [{"role": "user", "content": "Overview of §101 eligibility framework"}], "context_options": {"top_k": 5, "snippet_size": 2048}}}

// Step 2: Targeted gap fill via context retrieval (cheaper)
{"tool": "assistant_context", "arguments": {"query": "MPEP 2106.05(a) practical application technological improvement", "top_k": 3}}

assistant_context New Parameters

Multimodal and messages support expand what assistant_context can retrieve:

// With multimodal disabled (text-only retrieval, smaller response)
{
  "tool": "assistant_context",
  "arguments": {
    "query": "Section 101 eligibility",
    "top_k": 5,
    "multimodal": false,
    "include_binary_content": false
  }
}
// With messages input (multi-turn context retrieval)
{
  "tool": "assistant_context",
  "arguments": {
    "messages": [
      {"role": "user", "content": "What is the Alice/Mayo framework?"},
      {"role": "assistant", "content": "[prior response]"},
      {"role": "user", "content": "How does that apply to software patents?"}
    ],
    "top_k": 5
  }
}

Parameter notes:

  • query is now optional — either query OR messages is required
  • multimodal (optional bool): Enable image retrieval from PDFs. Only works with query input
  • include_binary_content (optional bool): Include base64 image data in response. Set false to reduce response size. Only works with query input
  • When using messages, the multimodal and include_binary_content parameters are ignored by the API

assistant_chat Response Format

assistant_chat returns clean text, not JSON. The response contains:

  1. The answer content
  2. A Sources section listing filename and a ~150-character excerpt per source
  3. A token summary line

Example response format:

[Answer text here...]

**Sources:**
- MPEP_Part2.md: "...the Alice/Mayo framework requires a two-step analysis. Step 2A asks whether..."
- Combined_Training_Materials.md: "...software claims directed to abstract ideas must provide significantly more..."

*Tokens: 35,205 prompt + 823 completion = 36,028 total*

🚀 Advanced Usage Patterns

Multi-Tool Workflow Integration

For complex patent prosecution tasks, combine the Pinecone Assistant with other USPTO MCPs:

Section 101 Rejection Response

# 1. Get Alice/Mayo framework guidance
assistant_strategic_multi_search_chat → domain: "section_101_eligibility"

# 2. Pull office action
PFW MCP → Download office action and specifications

# 3. Draft response with practical application arguments

Section 103 Obviousness Analysis

# 1. Research KSR and Graham factors
assistant_strategic_multi_search_chat → domain: "section_103_obviousness"

# 2. Analyze prior art
PFW MCP → Get prior art references and examiner analysis

# 3. Build secondary considerations case

PTAB Proceeding Preparation

# 1. Research institution standards
assistant_strategic_multi_search_chat → domain: "ptab_procedures"

# 2. Find similar proceedings
PTAB MCP → Search institution decisions

# 3. Analyze claim construction approach

Conversation Continuation Strategies

Use assistant_chat for follow-up questions after strategic searches:

{
  "tool": "assistant_chat",
  "arguments": {
    "messages": [
      {
        "role": "user",
        "content": "Based on the strategic search results, what specific KSR factors should I emphasize for a pharmaceutical formulation obviousness argument?"
      }
    ],
    "model": "gpt-4o",
    "include_highlights": true
  }
}

💡 Best Practices

When to Use Strategic Search vs Direct Chat

Use assistant_strategic_multi_search_chat when:

  • Starting research on a broad topic
  • Need comprehensive coverage of related issues
  • Want systematic analysis across multiple guidance documents
  • Preparing for complex prosecution or post-grant procedures

Use assistant_chat when:

  • Following up on specific findings
  • Need clarification on particular guidance
  • Have targeted questions about specific procedures
  • Want to explore nuances of a particular topic

Domain Selection Tips

  1. Match rejection type: Section 101 → section_101_eligibility, 103 → section_103_obviousness, etc.
  2. Consider technology: Software/AI → software_patents, Mechanical → mechanical_patents
  3. Legal issue focus: Use specific section domains (101, 102, 103, 112) for targeted analysis
  4. General research: Start with general_patent_law for broad questions
  5. Post-grant matters: Use ptab_procedures or claim_construction

Context Management

  • Strategic searches: Use for initial comprehensive research (20-30K tokens)
  • Follow-up chats: Use for specific questions (5-15K tokens)
  • Document specific: Switch to PFW/PTAB/FPD MCPs for specific documents
  • Monitor usage: Claude Desktop shows token consumption per interaction

🔍 Troubleshooting

Common Issues and Solutions

"Strategic search returns too broad results"

  • Use more specific query terms
  • Reduce max_searches parameter to 2
  • Switch to assistant_chat for targeted questions

"Need more specific technology guidance"

  • Use appropriate technology domain: software_patents or mechanical_patents
  • Follow up with specific section domains (101, 102, 103) for rejection-specific analysis
  • Cross-reference with PFW MCP for technology center guidance

"Context limit reached"

  • Switch to direct assistant_chat for remaining questions
  • Use PFW/PTAB/FPD MCPs for document-specific research
  • Save strategic search results and start new conversation

🔀 Multi-Assistant Workflows

Leverage up to 5 Pinecone Assistants (free tier) for specialized knowledge bases using the update_configuration tool.

Example: USPTO Research Setup

Create specialized assistants:

1. mpep-assistant          → MPEP 9th Edition core guidance (Combined_MPEP_9th_Edition_Part1-4.md)
2. updates-assistant       → Recent Federal Register notices & updates (Combined_MPEP_Updates.md)
3. case-law-assistant      → Federal Circuit & PTAB decisions (Not included)
4. training-assistant      → USPTO training materials & examples (Combined_Training_Materials.md)
5. international-assistant → PCT, Paris Convention, treaties (Not included)

Mid-Conversation Switching:

User: "What does MPEP § 2138.04 say about AI inventorship?"
[Using mpep-assistant - returns MPEP guidance]

User: use update_configuration(assistant_name: "updates-assistant")
✅ Switched to updates-assistant

User: "What recent Federal Register guidance addresses AI inventorship?"
[Now using updates-assistant - returns 2024 guidance]

User: use update_configuration(assistant_name: "case-law-assistant")
✅ Switched to case-law-assistant

User: "How have courts applied these principles in recent cases?"
[Now using case-law-assistant - returns Thaler v. Vidal analysis]

Benefits:

  • Single conversation spans multiple knowledge bases
  • 10 file limit per assistant (free tier) → 50 files total across 5 assistants
  • Specialized metadata and search patterns per assistant
  • Maintain conversation context while switching sources

Other Domain Examples

Medical Research:

  • clinical-trials → Clinical trial databases
  • literature → PubMed and research papers
  • guidelines → Medical society practice guidelines
  • pharmacology → Drug interactions and dosing
  • imaging → Radiology and diagnostic criteria

Financial Analysis:

  • sec-filings → 10-K, 10-Q, 8-K reports
  • earnings → Earnings call transcripts
  • analyst → Analyst reports and ratings
  • economic → Economic indicators and Fed minutes
  • regulations → GAAP, IFRS, regulatory guidance

Software Development:

  • documentation → API docs, technical specifications
  • architecture → System design patterns, best practices
  • security → Security guidelines, vulnerability reports
  • performance → Optimization guides, benchmarking data
  • frameworks → Framework-specific guides and examples

Multi-Assistant Best Practices

Setup Strategy:

  1. Domain Separation: Keep related but distinct content in separate assistants
  2. Size Management: Stay within 10 file limit per assistant by splitting logically
  3. Naming Convention: Use descriptive names that indicate content type
  4. Search Patterns: Customize strategic-searches.yaml per assistant's content

Workflow Optimization:

  1. Start Broad: Begin with general knowledge base (e.g., mpep-assistant)
  2. Narrow Down: Switch to specialized assistants for specific analysis
  3. Cross-Reference: Use multiple assistants to verify information
  4. Context Preservation: Claude Desktop maintains conversation context across switches

Token Management:

  • Use context tools first in each assistant to minimize costs
  • Switch assistants when you need different knowledge, not for token limits
  • Each assistant switch resets internal state but preserves Claude conversation

Example Multi-Assistant Research Flow:

// Step 1: General guidance
{
  "tool": "assistant_context",
  "arguments": {
    "query": "Section 101 software patent eligibility requirements"
  }
}

// Step 2: Switch to recent updates
{
  "tool": "update_configuration", 
  "arguments": {
    "assistant_name": "updates-assistant"
  }
}

// Step 3: Check recent guidance
{
  "tool": "assistant_context",
  "arguments": {
    "query": "AI patent eligibility guidance 2024 2025"
  }
}

// Step 4: Switch to case law
{
  "tool": "update_configuration",
  "arguments": {
    "assistant_name": "case-law-assistant" 
  }
}

// Step 5: Analyze precedent
{
  "tool": "assistant_strategic_multi_search_chat",
  "arguments": {
    "query": "software patent eligibility Federal Circuit decisions",
    "domain": "section_101_eligibility"
  }
}

This approach gives you comprehensive coverage while staying within free tier limits and optimizing token usage across multiple specialized knowledge bases.

📊 Answer Evaluation

evaluate_answer (Paid Plan Only)

Evaluate AI answer quality against a ground truth.

Requires Pinecone Standard (paid) plan. Free tier users will receive an error. Rate limited to 20 requests/minute.

Use case: After getting an answer from assistant_chat, evaluate its accuracy against a known-correct answer.

{
  "tool": "evaluate_answer",
  "arguments": {
    "question": "What is the Alice/Mayo two-step test?",
    "answer": "The Alice/Mayo framework has two steps. Step 2A asks whether the claim is directed to a judicial exception (abstract idea, law of nature, or natural phenomenon). Step 2B determines whether additional elements amount to significantly more than the judicial exception.",
    "ground_truth_answer": "The Alice/Mayo test is a two-step framework established by Alice Corp. v. CLS Bank (2014) and Mayo Collaborative Services v. Prometheus (2012). Step 1 asks whether the claim is directed to a judicial exception. Step 2 asks whether the claim includes additional elements that amount to significantly more than the exception alone, providing an inventive concept."
  }
}

Response format:

**Alignment Scores:**
- Alignment (overall): 0.847
- Correctness (precision): 0.900
- Completeness (recall): 0.800

**Evaluated Facts:**
  [entailed] Two-step framework
  [entailed] Step 2A asks whether claim is directed to judicial exception
  [entailed] Judicial exceptions include abstract ideas, laws of nature, natural phenomena
  [neutral] Established by Alice Corp. v. CLS Bank (2014)
  [contradicted] Step 2 is called Step 2B (answer uses "Step 2B" for what ground truth calls "Step 2")

*Tokens: 1,204 prompt + 87 completion = 1,291 total*

Score interpretation:

  • alignment >= 0.8: High quality answer
  • alignment 0.5–0.8: Acceptable but incomplete
  • alignment < 0.5: Significant gaps or errors

Workflow example — evaluate after delegation:

// Step 1: Get answer via delegation
{
  "tool": "assistant_chat",
  "arguments": {
    "messages": [{"role": "user", "content": "What is the Alice/Mayo two-step test?"}],
    "model": "gpt-4o"
  }
}

// Step 2: Evaluate the answer against a known-correct reference
{
  "tool": "evaluate_answer",
  "arguments": {
    "question": "What is the Alice/Mayo two-step test?",
    "answer": "[paste the answer from Step 1]",
    "ground_truth_answer": "[your verified reference answer]"
  }
}

This document is maintained alongside the Pinecone Assistant MCP. For the latest version and updates, see the main repository.