AI Application Architecture

Write a story of about 400 words. A bus driver in Chattogram finds a lost bag on the last trip of the night. Start the story right away, with no introduction.

Explain each of these in one simple sentence: SSE (Server-Sent Events), WebSockets, and HTTP chunked transfer.
Then pick one for each feature and give one reason.
Feature 1: An AI chat answer appears word by word. Data only goes from server to browser. The route is a normal HTTP proxy.
Feature 2: A voice call where the user and the AI can talk at the same time.
Feature 3: A fallback for an old company proxy that removes the SSE headers.
Then list one cost of WebSockets compared to SSE.

Act as the API server of a research agent. Reply only with an HTTP status line and a JSON body. No other text.
Rules:
- POST /jobs starts a job. Reply at once with 202 and a JSON body with job_id and status "queued". The first job_id is job_001. The job takes 245 seconds.
- GET /jobs/{job_id} gives the status. Statuses are queued, running, done, and failed.
- I tell you the elapsed seconds in each GET request. Use this table:
  0 to 59 seconds: running, step "1 search sources"
  60 to 179 seconds: running, step "2 read 12 sources"
  180 to 244 seconds: running, step "3 write the report"
  245 seconds or more: done, with a result field that has a short fake report title
- When running, also show progress as a percent: elapsed divided by 245, rounded to a whole number.
Say READY if you understand.

POST /jobs
{"topic": "Mobile payment use in rural Bangladesh"}

GET /jobs/job_001
elapsed: 40 seconds

GET /jobs/job_001
elapsed: 120 seconds

GET /jobs/job_001
elapsed: 245 seconds

Act as an API server for report generation. Each new run costs 0.40 USD. Reply only with JSON for each request I send.
Rules:
- Every request has an idempotency_key and a body.
- A new key makes a new run with a new run_id (run_1, run_2, ...). Reply with charged: 0.40 and replayed: false.
- The same key with the same body returns the stored response again, with replayed: true and charged: 0. No new run.
- The same key with a different body returns the error code 422 with a short message.
Say READY if you understand.

Handle these three requests in order. After the third, add a line with the total USD charged.
1. idempotency_key: k-1001, body: {"report": "monthly sales"}
2. idempotency_key: k-1001, body: {"report": "monthly sales"} (the user refreshed the page)
3. idempotency_key: k-1001, body: {"report": "yearly sales"}

Now a long run. Run run_77 has 5 steps. Each step costs 0.10 USD, and a failed step costs nothing. Steps 1, 2, and 3 finished and were already charged. The server saved a checkpoint after each one. Step 4 failed with a network error. The client retries with the same idempotency_key.
1. Which steps do you run now, and what do you charge?
2. What would the retry cost if there were no checkpoints?
3. Name one thing the checkpoint must contain.

You can use one tool: get_weather(city: string), which returns live weather.
User question: Do I need an umbrella in Sylhet today?
Reply ONLY with the tool call as JSON, in the form {"tool": ..., "arguments": {...}}. Do not answer the question yet.

You have no tools and no live data. Answer from general knowledge only. Say clearly that you cannot check today's weather, and give the most useful advice you can.
User question: Do I need an umbrella in Sylhet today?

Act as an LLM gateway. The fallback chain is: Provider 1 (has tools), then Provider 2 (has tools), then Provider 3 (no tools, so use a degraded prompt that says: you have no tools, answer from general knowledge).
Rules: on a 503 or 429 error, go to the next provider. On success, stop. Record the provider that answered, the number of attempts, and the cost.
Costs per call: Provider 1 0.010 USD, Provider 2 0.006 USD, Provider 3 0.002 USD. Failed calls cost 0.
Three requests came in:
Request X: Provider 1 returns 200.
Request Y: Provider 1 returns 503, Provider 2 returns 200.
Request Z: Provider 1 returns 503, Provider 2 returns 429, Provider 3 returns 200.
Give a table with: request, provider used, attempts, prompt used (normal or degraded), cost. Then add the total cost.