Dependency Injection for AI Components
Register .NET AI clients, tools, prompts, and retrieval services with explicit lifetimes, safe request state, and clear disposal ownership.
Ai-Engineering
Treat the model finish reason as a control signal
Check why generation stopped before accepting model output, and give incomplete responses an explicit application outcome.
Persist ingestion checkpoints at source boundaries
Track progress per consumer and source ordering boundary, and advance each checkpoint only after the corresponding ingestion work succeeds.
Treat provider throttling as a shared capacity signal
Let one throttling response reduce pressure from every caller sharing the affected capacity pool.
Human Approval as a System Boundary
Design human approval around an exact operation, with reviewable payloads, durable decisions, expiry, and execution checks that survive retries.
Store a source checksum to skip unchanged documents
Compare a document’s source checksum, metadata, and ingestion profile before repeating extraction, chunking, embedding, and index writes.
Keep tool results inside a token budget
Bound every model-visible tool result before it enters the next prompt.
Your First Qdrant Collection
Run Qdrant locally, connect with the official .NET client, and create a collection whose vector size and distance match the embedding model.
Run production evaluations outside the user response path
Hand production samples to a separate evaluation worker instead of making users wait for an evaluator model.
AI Systems Need Runtime Budgets
How to bound time, tokens, retries, tool calls, and estimated cost across a complete AI execution instead of limiting each call in isolation.