Function Call Context

Provide function call context as part of the conversation history when starting a new Voice Agent session.
Voice Agent

When starting a new conversation session with the Voice Agent, you can provide historical context about previous function calls using the agent.context parameter. This allows the agent to maintain awareness of past function executions, enabling more coherent and contextual conversations.

Purpose

Function call context is particularly useful when:

  • Resuming a conversation from a previous session
  • Providing background context about function calls already executed
  • Maintaining continuity in multi-session interactions
  • Allowing the agent to reference previous function results

Message Schema

Function call context is included in the conversation history using the History message type with a function_calls array. Each function call entry contains complete information about the execution, including arguments and results.

Function call context can be disabled by setting settings.flags.history to false in the agent configuration. History is enabled by default.

JSON
1{
2 "type": "Settings",
3 "settings": {
4 "flags": {
5 "history": true // true by default set to false to disable
6 }
7 },
8 "agent": {
9 "context": {
10 "messages": [
11 {
12 "type": "History",
13 "role": "user",
14 "content": "What's the weather like in New York?"
15 },
16 {
17 "type": "History",
18 "function_calls": [
19 {
20 "id": "fc_weather_12345",
21 "name": "get_weather",
22 "client_side": true,
23 "arguments": "{\"location\": \"New York\"}",
24 "response": "The current weather in New York is partly cloudy with a temperature of 295.15°K."
25 }
26 ]
27 },
28 {
29 "type": "History",
30 "role": "assistant",
31 "content": "The weather in New York is partly cloudy with a temperature of about 72°F (295.15°K)."
32 }
33 ]
34 },
35 "think": {
36 "provider": {
37 "type": "open_ai",
38 "model": "gpt-4o-mini"
39 },
40 "functions": [
41 {
42 "name": "get_weather",
43 "description": "Get the current weather for a specific location",
44 "parameters": {
45 "type": "object",
46 "properties": {
47 "location": {
48 "type": "string",
49 "description": "The city or location to get weather for"
50 }
51 },
52 "required": ["location"]
53 }
54 }
55 ]
56 }
57 }
58}

Multiple Message Types

Function call history works seamlessly with Conversation Context to provide complete context. You can mix conversation messages and function call messages in the same context array.

Function Call Context History Structure

Each function call in the history includes the following fields:

FieldTypeDescription
idStringA unique identifier for the function call
nameStringThe name of the function that was called
client_sideBooleanIndicates if the function was executed client-side or server-side
argumentsStringJSON string containing the arguments passed to the function
responseStringThe response/result returned by the function

Benefits of Function Call Context History

Including function call history in your agent context provides several advantages:

Continuity

The agent understands what functions were previously called and their results, enabling natural conversation flow across sessions.

Efficiency

The agent can reference previous function results instead of making redundant calls for the same information.

Context Awareness

The agent can make informed decisions based on the complete interaction history, including both conversations and actions taken.

Natural Interactions

Users can refer to previous actions (“like we did before”) and the agent will understand the context.

Best Practices

Unique IDs

In most cases the LLM will assign ids. If the LLM does not assign ids, Deepgram will assign ids to the function calls.

Complete Information

Include both the arguments passed to the function and the complete response received.

Chronological Order

Maintain the chronological order of function calls and conversations in your history.

Relevant Context Only

Include only the function call history that’s relevant to the current conversation to avoid overwhelming the context.