BuildSpark AI
Dismiss

BuildSpark AI: Understand AI. Make it work for you. New here? Start here.

Home · Builder path

Builder path

Pick what you're building. We'll give you the cards in order.

A card is a short page that explains one idea and shows it working. The basics come first.

Know the basics already? Skip to step 2: pick what you're building

Step 1: the basics

Seven cards, in order.

  1. 1 Tokens
  2. 2 Context Window
  3. 3 Prompting
  4. 4 Prompt Caching
  5. 5 Model Selection & Routing
  6. 6 Hallucination
  7. 7 Evaluation basics

Step 2: pick what you're building

Answer from your own documents

Split your files, find the right passage, and tie the reply to it.

  1. ChunkingSplit pages so search returns a piece that still names its subject.
  2. Embeddings for retrievalFind passages by meaning, and check search returns the one you'd pick.
  3. RAGRAG (retrieval-augmented generation): answer from your own documents, not memory.
  4. Citation & grounding checkProve each answer sentence is in the source it cites.

Start with Chunking →

When you finish, you can split your files, find the right passage, and tie the reply to it. Next: Build an agent.

Build an agent

Let AI models call your tools, run each task to a clear stop, and limit what a document or tool result can make the agent do.

  1. Structured OutputsRequest the fields you actually read, so code can read them.
  2. Tool callingLet the model fetch facts by calling a function in your code.
  3. Agent loopRun a tool-using task to a clear stop, and see what each tool returned.
  4. Tool design/MCPDescribe each tool so the model calls the right one with the right inputs.
  5. Context engineeringSend each step only what it needs, and stop paying for the rest.
  6. Prompt injectionLimit what a document or tool result can make your agent do.

Start with Structured Outputs →

When you finish, you can let AI models call your tools, run each task to a clear stop, and limit what a document or tool result can make the agent do. Next: Ship it to production.

Ship it to production

Catch a broken change before users do, see why a request failed, check cost, speed, privacy and the model's license, and spot quality drops after launch.

  1. Production evaluationCatch a broken change before users do, with saved real examples.
  2. ObservabilitySee why a request failed: the question, what search and tools returned, the reply.
  3. Cost forecastingPrice the workflow from the published rates and how often it runs.
  4. LatencySee which step the user waits on.
  5. Fallbacks & rate limitsRetry, switch or stop cleanly when a model errors or hits a rate limit.
  6. Privacy & data handlingSend only what the question needs, and know what the provider keeps.
  7. Monitoring & driftCatch quality drops when the model or your documents change.
  8. LicensingConfirm you may ship this model to users, and what credit it needs.

Start with Production evaluation →

When you finish, you can catch a broken change before users do, see why a request failed, check cost, speed, privacy and the model's license, and spot quality drops after launch. Next: Safety: test that your product answers what it should and refuses the rest.