Skip to content

README.md

Latest commit

 

History

History
224 lines (161 loc) · 8.89 KB

README.md

File metadata and controls

224 lines (161 loc) · 8.89 KB

Pinecone Assistant MCP - Claude Skills

Three Claude skills are included in this directory for guided research workflows in Claude Code and Claude.ai. Skills provide Claude with domain-specific instructions for tool selection, token optimization, and research patterns — without requiring you to explain the workflow each session.

Available Skills

Skill Directory Audience Primary Tool
pinecone-assistant skills/pinecone-assistant/ Any corpus assistant_context
pinecone-assistant-uspto skills/pinecone-assistant-uspto/ USPTO / MPEP Domain-specific search
pinecone-assistant-paid-plan skills/pinecone-assistant-paid-plan/ Paid plan / agentic assistant_chat

Installation

Claude Code (Claude CLI)

Copy the skill directory into your Claude skills folder:

# Copy individual skills (run from the pinecone_assistant_mcp directory)
cp -r skills/pinecone-assistant ~/.claude/skills/
cp -r skills/pinecone-assistant-uspto ~/.claude/skills/
cp -r skills/pinecone-assistant-paid-plan ~/.claude/skills/

Windows (PowerShell):

# From the pinecone_assistant_mcp directory
Copy-Item -Recurse skills\pinecone-assistant $env:USERPROFILE\.claude\skills\
Copy-Item -Recurse skills\pinecone-assistant-uspto $env:USERPROFILE\.claude\skills\
Copy-Item -Recurse skills\pinecone-assistant-paid-plan $env:USERPROFILE\.claude\skills\

After copying, invoke the skill in Claude Code:

/pinecone-assistant

Claude.ai (Web Interface)

Upload the SKILL.md file for the skill you want using the Claude.ai Projects feature or attach it as context at the start of a conversation.

Skill Descriptions

pinecone-assistant — Generic Corpus Research

File: skills/pinecone-assistant/SKILL.md

Best for: Any document corpus — legal, medical, technical, financial, or custom

What it teaches Claude:

  • Tool selection decision tree (which tool to use and when)
  • Token budget management for free-tier users
  • Multi-assistant setup and switching with update_configuration
  • Query formulation best practices
  • Parameter reference for all 6 tools

Tool priority (cost-optimized):

  1. assistant_context — single-topic retrieval (default first choice)
  2. assistant_strategic_multi_search_context — multi-angle raw retrieval
  3. assistant_strategic_multi_search_chat — AI synthesis across search patterns
  4. assistant_chat — delegated synthesis (use sparingly on free tier)

Key guidance:

  • Always try assistant_context before escalating to AI tools
  • Stateless by default: only include conversation history when questions explicitly reference prior answers
  • top_k=5, snippet_size=2048 is the recommended default for assistant_context

Triggers: Generic document research, knowledge base lookup, tool selection questions

pinecone-assistant-uspto — USPTO / MPEP Research

File: skills/pinecone-assistant-uspto/SKILL.md

Best for: Patent law research using the included USPTO corpus (MPEP, examination guidelines, training materials)

What it teaches Claude (extends pinecone-assistant):

  • 11 pre-configured search domains with descriptions and use cases
  • Domain selection guide mapping legal issue types to domains
  • Core research workflows for §101, §103, §112, and PTAB proceedings
  • Query patterns with MPEP section numbers and examination terminology
  • Cross-MCP integration with PFW, FPD, and PTAB servers

USPTO Search Domains:

Domain Use When
section_101_eligibility Patent eligibility (Alice, Mayo) rejections
section_103_obviousness Obviousness rejections and combination of references
section_112_requirements Written description, enablement, indefiniteness
section_102_novelty Anticipation and novelty rejections
claim_construction Claim interpretation and BRI
ptab_procedures IPR, PGR, PTAB practice and procedure
mechanical_patents Mechanical and structural inventions
software_patents Software, algorithms, and computer-implemented inventions
general_patent_law Broad patent law questions not fitting specific domains

Key workflows:

  • Office Action Response: Domain search → MPEP guidance → prosecution strategy
  • §101 Eligibility Analysis: Alice/Mayo two-step framework research
  • §103 Obviousness: TSM test, motivation to combine, objective indicia
  • PTAB Research: IPR/PGR filing requirements, estoppel, trial procedures

Triggers: Patent law questions, MPEP lookups, office action research, USPTO examination procedure

pinecone-assistant-paid-plan — Agentic Delegation + Answer Evaluation

File: skills/pinecone-assistant-paid-plan/SKILL.md

Best for: Users on a paid Pinecone plan, or agentic workflows where Claude's context window is the primary constraint

What it teaches Claude (extends pinecone-assistant):

  • When to use assistant_chat as a sub-agent for delegation
  • How to use evaluate_answer to score answer correctness, completeness, and alignment against a ground truth
  • Iterative refinement loop: generate → evaluate → improve → re-evaluate
  • RAG evaluation workflows for batch testing question/answer pairs
  • context_options tuning for controlling Pinecone's internal retrieval
  • Agentic chaining patterns for multi-part research
  • Mixing delegation with context retrieval for cost efficiency
  • Model selection guidance for different research tasks

Primary tools: assistant_chat, evaluate_answer

Why delegation preserves context window:

  • Claude receives only the synthesized answer (~500–2000 tokens)
  • Instead of raw document chunks (which can be 10–32K tokens per query)
  • Pinecone handles retrieval and synthesis internally

Answer evaluation:

evaluate_answer scores a generated answer against a known ground truth:

Metric What it measures
correctness Precision — how much of the answer is factually correct
completeness Recall — how much of the ground truth the answer covers
alignment Harmonic mean of correctness + completeness (F1 score)

Per-fact entailment breakdown: entailed ✅ / contradicted ❌ / neutral ➖

context_options tuning:

Setting Tokens to LLM Use When
top_k=3, snippet_size=1024 ~3K Focused single-concept questions
top_k=5, snippet_size=2048 ~10K Standard research (recommended)
top_k=10, snippet_size=2048 ~20K Broad topic coverage
top_k=16, snippet_size=2048 ~32K Maximum coverage

Agentic chaining pattern:

# Chain independent calls without history — each is stateless
result_a = assistant_chat("First research question")
result_b = assistant_chat("Second research question")
result_c = assistant_chat("Third research question")

# Claude synthesizes the three compact answers

Triggers: "delegate to the assistant", "let Pinecone synthesize", "agentic research", "evaluate answer quality", "validate the answer", "check correctness", "score my answer", "context window is full", "chain multiple research tasks"

Choosing the Right Skill

Am I using the USPTO / MPEP corpus?
├── Yes → pinecone-assistant-uspto
│   (extends base skill with domain selection and patent law workflows)
│
└── No (or any corpus)
    ├── On a paid plan or agentic workflow?
    │   └── Yes → pinecone-assistant-paid-plan
    │       (assistant_chat delegation as primary tool)
    │
    └── On free tier or general use
        └── pinecone-assistant
            (context-only tools first, cost-optimized)

You can install all three skills — Claude will use the most specific one that matches the context.

Relationship to Prompt Templates

Skills and prompt templates complement each other:

  • Prompt templates (in the Claude + menu) provide structured workflows for common research tasks
  • Skills provide Claude with deeper guidance on tool selection, parameters, and optimization
Prompt Recommended Skill
deep_research pinecone-assistant or pinecone-assistant-uspto
quick_lookup pinecone-assistant
comparative_research pinecone-assistant
delegated_research pinecone-assistant-paid-plan

See PROMPTS.md for full prompt template documentation.

Updating Skills

Skills are plain Markdown files. To update a skill, edit the SKILL.md file and copy it again to ~/.claude/skills/. Changes take effect in the next Claude Code session.

To verify a skill is loaded in Claude Code:

/skills

For MCP server setup, see README.md and INSTALL.md.

For prompt template documentation, see PROMPTS.md.