LLM-OPS - Setup Guide¶
Prerequisites¶
- Node.js 18+ (for npx)
- Any supported agent: Claude Code, Antigravity, Cursor, Gemini CLI, or Codex CLI
- Basic understanding of LLM concepts (RAG, embeddings, vector DBs)
Capabilities¶
| Capability | Description |
|---|---|
| RAG Implementation | Build retrieval-augmented generation pipelines |
| Embeddings | Create and manage embeddings with Pinecone, Chroma, pgvector |
| Vector Databases | Select and configure the right vector store for your use case |
| Fine-Tuning | Fine-tuning workflows for domain-specific models |
| Prompt Engineering | Advanced prompt design patterns |
| Cost Management | Track and reduce LLM API costs |
| Quality Evaluations | Eval frameworks for model output quality |
| Semantic Caching | Cache LLM responses to reduce latency and cost |
| Streaming | Real-time streaming response architectures |
| Agent Architectures | Production AI agent design patterns |
Installation¶
npx skills add sickn33/antigravity-awesome-skills/llm-ops
Quick Start¶
Trigger the skill when you need specialized LLM operations assistance:
- "Set up a RAG pipeline for our documentation"
- "Compare Pinecone vs Chroma for our use case"
- "Optimize our LLM costs - we're spending too much on API calls"
- "Design an eval framework for our agent responses"
- "Implement semantic caching for repeated queries"
Key Patterns¶
RAG Pipeline Architecture¶
Documents → Chunking → Embeddings → Vector Store → Retrieval → LLM → Response
Cost Optimization Strategies¶
| Strategy | Savings | Complexity |
|---|---|---|
| Semantic caching | 30-60% | Low |
| Prompt compression | 20-40% | Medium |
| Model routing (cheap model for easy queries) | 40-70% | Medium |
| Batch processing | 50% | Low (if latency-tolerant) |
| Fine-tuned small models | 60-80% | High |
Vector Database Selection¶
| Database | Best For | Limitations |
|---|---|---|
| Pinecone | Managed, zero-ops, fast | Cost at scale, vendor lock-in |
| Chroma | Local dev, open source | Not production-hardened |
| pgvector | Postgres-native, SQL queries | Slower than purpose-built |
| Weaviate | Hybrid search, GraphQL | Complex setup |
| Qdrant | Rust performance, filtering | Smaller ecosystem |
CorpusIQ Use Cases¶
-
Agent Knowledge Base: Build RAG pipelines over the CorpusIQ docs, Hermes operating protocols, and governance rules - agents query their own knowledge base instead of relying on training data.
-
Cost Tracking: Monitor and optimize LLM spend across all CorpusIQ-connected agents (Sonnet, Opus, DeepSeek, Qwen). Route queries to the cheapest capable model.
-
Quality Evaluations: Implement eval suites for agent responses - measure accuracy, helpfulness, and safety before deploying new agent versions.
-
Semantic Caching: Cache frequent agent queries (e.g., "What's the preflight gate?") to reduce API costs by 30-60%.
Troubleshooting¶
| Symptom | Likely Cause | Fix |
|---|---|---|
| Embeddings dimension mismatch | Wrong model selected | Match embedding dimension to vector DB config |
| RAG returns irrelevant chunks | Chunk size too large/small | Tune chunk size (500-1000 tokens optimal) |
| Cost still high after caching | Cache miss rate too high | Widen cache key matching (semantic, not exact) |
| Fine-tuning job fails | Insufficient training data | Minimum 100 examples for LoRA fine-tuning |
← Skills Catalog | Marketplace → ↑ Skills Home
Curated by CorpusIQ - one MCP endpoint, all your business tools. Content remains attributed to original authors and repositories.
Integration with CorpusIQ¶
CorpusIQ is the connector layer that lets AI assistants read the systems a business already runs on - accounting, payments, CRM, email, files and analytics - and answer with the source named. It is read-only by design: it retrieves data and never writes to, or acts in, a connected system, which is what makes it safe to point at live accounts.
This page is part of the CorpusIQ directory, the reference we maintain for the MCP server and agent-skill ecosystem. Run alongside CorpusIQ, a third-party server like this one can be combined with a company's own business data in a single answer.