AI Best Practices

Workflow Automation & AI Tools

AI Models & Frameworks

Learning & Understanding

General Best Practices

Token spend & FinOps control

Treat model usage like cloud spend: see who burns tokens, set budgets, and route traffic through one metered gateway—FinOps discipline applied to LLM inference instead of only to VMs and buckets.

My Document pipeline

One practical PDF-to-knowledge flow: normalize PDFs locally, archive and OCR in Paperless, enrich with Paperless-AI, optionally run summarization and translation through durable Temporal workflows, then work with the corpus in AnythingLLM (with Paperless-ngx wired in for quick checks). Finally, load the consolidated knowledge into OpenRAG and expose it through a home-grown MCP tool so every other agent or chat client can reuse the same retrieval surface.

1. 
   Stirling PDF Locally split, merge, rotate, or clean PDFs before ingest so uploads are consistent and smaller.
2. 
   Paperless-ngx Upload the prepared PDFs; Paperless-ngx stores, indexes, and runs document consumption (tags, correspondents, searchable archive).
3. 
   Paperless-AI Let Paperless-AI process documents after ingest (classification, metadata, RAG-oriented workflows, depending on your setup); wait until processing finishes before relying on extracted text or tags.
4. 
   Temporal Orchestrate a follow-on workflow: summarize long documents, translate content into other languages, chain human-in-the-loop steps, and retry activities reliably. Temporal keeps state and timers for you so these pipelines survive restarts and remain observable. temporalio/temporal is the open-source server; pair it with workers that call your LLM or translation APIs.
5. 
   AnythingLLM (Docker) Use a self-hosted AnythingLLM instance (for example via Docker) to pull processed content into a private chat/RAG workspace. As Paperless-ngx is integrated with AnythingLLM, you can browse and question the same document set from the chat UI and quickly verify that OCR, tags, and downstream summaries or translations look right. The linked guide covers storage mounts, environment variables, and reaching host services from the container.
6. 
   OpenRAG Push the full, vetted corpus into OpenRAG—the upstream stack combines Langflow workflows, Docling parsing, and OpenSearch-backed retrieval so you get a single package for ingestion, search, and chat over your documents. From there, connect a custom MCP server you own (wrap OpenRAG’s APIs or follow the same Model Context Protocol patterns as the project’s openrag-mcp helper) and register it in Cursor, Claude Desktop, or any other MCP-aware host. Later, every other AI agent or assistant can call those tools instead of re-implementing access to your knowledge base.