Skip to content

sampling.md

Latest commit

 

History

History
64 lines (44 loc) · 1.41 KB

sampling.md

File metadata and controls

64 lines (44 loc) · 1.41 KB

Sampling Lab ⚗️

Generate model replies locally (echo) or with OpenAI – the switch is automatic.

Environment Manager Used
OPENAI_API_KEY unset sampling.DefaultManager (echo)
OPENAI_API_KEY set OpenAISamplingManager (GPT‑4o)

1  Synchronous Reply

curl -s -X POST :8080/rpc \
  -d '{"jsonrpc":"2.0","id":1,"method":"sampling/createMessage","params":{"systemPrompt":"Tell a joke."}}' | jq .samplingMessage.content.text

2  Streaming Reply

Open two terminals.

Terminal 1 – listen to SSE

curl -sN localhost:8080/events | jq -c

Terminal 2 – fire the request

curl -s -X POST :8080/rpc \
  -d '{"jsonrpc":"2.0","id":2,"method":"sampling/createMessageStream","params":{"systemPrompt":"Write a haiku about cheese."}}'

The first token comes back through RPC, the rest arrives over the SSE stream.

3  Custom Preferences

curl -s -X POST :8080/rpc -d '{
  "jsonrpc":"2.0","id":3,
  "method":"sampling/createMessage",
  "params":{
    "systemPrompt":"You are a pirate.",
    "temperature":1.1,
    "maxTokens":64,
    "stopSequences":["Arrr"]
  }}' | jq .samplingMessage.content.text

What Next?

  • Combine everything: prompt → embed resource → sampling → stream.
  • Implement a custom ToolExecutor so GPT can call your own Go functions.

Have fun exploring! 🧪