cloudvoyant/llm

llm · experimental

Context skill for LLM engineering: agent frameworks, document/image processing, open-weight model serving, retrieval-augmented generation, and evaluation tooling.

When to load recipes

| You are working on... | Load recipe | |---|---| | Agent with Vercel AI SDK or Anthropic SDK | references/recipes/agents-ai-sdk.md | | Agent with LangChain or LangGraph | references/recipes/agents-langgraph.md | | Tool calling in any SDK | references/recipes/tool-calling.md | | PDF text extraction, OCR, or document chunking | references/recipes/document-processing.md | | Image tiling for Vision APIs or PDF rasterization | references/recipes/image-tiling.md | | Serving open-weight models on AWS | references/recipes/serving-aws.md | | Serving open-weight models on GCP | references/recipes/serving-gcp.md | | RAG pipeline on AWS (OpenSearch, Bedrock) | references/recipes/rag-aws.md | | RAG pipeline on GCP (Vertex AI, BigQuery) | references/recipes/rag-gcp.md | | RAG with OSS vector stores (Qdrant, Chroma) | references/recipes/rag-oss.md | | LLM evaluation, RAGAS, LangSmith, regression testing | references/recipes/llm-eval.md |

Philosophy

LLM engineering is systems engineering with a probabilistic core. The model is just one component in a pipeline that includes document ingestion, embedding, retrieval, prompt construction, tool execution, and output evaluation. These recipes treat each component as an independently testable, observable subsystem. Agent loops have explicit state machines. RAG pipelines have measurable retrieval quality. Serving infrastructure has latency budgets and cost models. Eval is not optional — it is the feedback loop that makes everything else improvable.

Recipes

• Agents with Vercel AI SDK / Anthropic SDK — agent loop, tool calling, streaming, multi-turn, human-in-the-loop
• Agents with LangChain/LangGraph — chains, state graphs, conditional edges, memory, checkpointing
• Tool Calling — defining schemas, parallel tool use, error recovery, tool result handling
• PDF and Document Ingestion — text extraction, OCR, chunking strategies (recursive, semantic)
• Image Processing for Multimodal LLMs — PDF rasterization, tiling large images for Vision APIs, canvas/WebGL rendering
• Serving on AWS — EC2 with vLLM, SageMaker endpoints, spot instances, quantization
• Serving on GCP — Vertex AI Model Garden, Cloud Run with vLLM, Compute Engine GPU
• RAG on AWS — OpenSearch vector store, Bedrock embeddings, S3 + Lambda pipeline
• RAG on GCP — Vertex AI Search, BigQuery vector search, RAG Engine
• RAG with OSS — Qdrant, Chroma, hybrid search (BM25 + vector), RAGAS evaluation
• LLM Eval Tooling — RAGAS, LangSmith tracing, custom eval harnesses, regression tracking

References

• Vercel AI SDK
• Anthropic SDK
• LangChain
• LangGraph
• vLLM
• Qdrant
• Chroma
• RAGAS
• LangSmith