Ch 7: Sampling & Elicitation

Ch 7 — Sampling & Elicitation Under the Hood

Wire formats, model preferences, includeContext, elicitation schemas, capability negotiation

Index ← High Level

Under the Hood

Click play or press Space to begin...

Step- / 10

ASampling Request Wire FormatThe full createMessage request structure

send

createMessagemessages, model
prefs, maxTokens

params

tune

modelPreferenceshints, cost, speed,
intelligence priority

arrow_downward includeContext: requesting conversation history

BincludeContextServer can request the host include conversation context

history

noneOnly server's
messages sent

chat_bubble

thisServerInclude context
from this server

forum

allServersInclude full
conversation

arrow_downward Sampling response: what comes back

CSampling ResponseThe result returned to the server

psychology

LLM GeneratesHost sends to LLM
gets completion

returns

chat

Responserole, content,
model used

includes

stop_circle

stopReasonendTurn, maxTokens,
or stopSequence

arrow_downward Elicitation request: JSON Schema for user input

DElicitation RequestDefining what the server needs from the user

send

elicitation/createmessage +
requestedSchema

schema

JSON Schemastring, number,
boolean, enum

response

task_alt

Action + Contentaccept, decline,
or dismiss

arrow_downward Capability negotiation for sampling and elicitation

ECapability NegotiationBoth sides must opt in during initialization

dns

Server DeclaresI need sampling
and/or elicitation

handshake

hub

Host DeclaresI support sampling
and/or elicitation

result

handshake

AgreedBoth sides know
what's available

arrow_downward Agentic loop: combining sampling with tools

FAgentic Loop PatternServer drives a multi-step workflow using sampling + tools

psychology

SampleAsk LLM to
analyze / decide

act

build

ExecuteServer acts on
LLM's response

loop

RepeatUntil task
is complete