LLMs And Conversation Behavior In DialNexa

Compare DialNexa language models and tune prompts, temperature, fallback behavior, preprocessing, functions, and response quality.

LLMs in DialNexa are the part of a Voice AI call that decides what to say or do next. The selected model reads the prompt, conversation history, dynamic variables, knowledge context, and function definitions, then returns a reply or an action. If the transcript is correct but the answer is wrong, this is where the investigation usually starts.

The model is not a mind reader. If the outcome matters, write the instruction and define the field. Vibes are not a configuration format.

DialNexa LLMs And Conversation Behavior

Your workspace may show a subset of model families depending on what is enabled.

Model family	Use it when	Validate before production
OpenAI	You want the safest general-purpose default for prompts, functions, structured outputs, summaries, and post-call extraction.	Function arguments, response format, refusal wording, and behavior on long conversation history.
Google and Gemini	You want to compare reasoning, long-context behavior, or direct realtime speech behavior where Gemini Speech to Speech is enabled.	Tool calling, summary consistency, strict instruction following, interruptions, and first audio timing.
Groq	You need fast model responses or a low-latency fallback where enabled.	Response length, function behavior, and whether speed still leaves enough reasoning quality.

For help comparing language models, see the voice AI provider selection guide. For business-system actions the model can trigger or prepare, start with using integrations in agents.

What The Model Selector Does

Behavior	What users should know
Default model	New agents try to select GPT-4o Mini when it is available, otherwise the first available non-deleted model.
Pricing preview	The model selector can show ₹x.xx/min beside each model.
Fallback pricing preview	Fallback model choices can show ₹0.00/min while fallback billing rates are being wired. Treat the dashboard and billing transactions as the source of truth for actual charges.
Published state	Published versions can disable model changes. Edit a draft when comparing models.
Provider logos	The selector visually distinguishes OpenAI, Google, and Groq model families.
Settings button	The settings popover controls temperature, fallback LLM, fallback delay, predictive preprocessing, and Prompt Caching where available.
Current Time Awareness	The timezone control tells the agent what local time to use for “today,” “tomorrow,” business hours, and scheduling windows.

Speech To Speech Models

Speech to Speech models are different from cascaded text LLMs. They listen and speak directly, so the dashboard hides separate transcriber, voice model, and Audio Cache controls.

Model path	What to test
OpenAI realtime	Interruption handling, function behavior, welcome timing, and whether the voice fits the caller.
Gemini 3.1 Flash Live	Gemini voice fit, automatic activity detection, tool calls, voicemail detection, welcome startup behavior, long-call continuity, and the visible INR per-minute pricing preview.

Use Speech to Speech for latency-sensitive calls where separate control over STT and TTS is less important than natural turn taking. Use cascaded agents when you need more control over the transcriber, voice provider, fallback STT, Audio Cache, and separate provider costs.

Temperature

The LLM Temperature slider runs from 0 to 1 in the dashboard. Lower values are better for function calls and structured results.

Call type	Suggested direction
Booking, payments, eligibility, compliance, or structured extraction.	Keep temperature low. Stable arguments matter more than colorful phrasing.
Support intake or objection handling.	Start low, then test a moderate value only if responses are too stiff.
Knowledge-heavy calls.	Keep temperature low until retrieval and answer quality are proven.
Regulated scripts.	Keep temperature low and write explicit allowed and disallowed behavior.

Fallback LLM

Fallback LLM is a per-agent setting that can start a backup model after a configured delay. New default agent configuration enables fallback where available, uses Gemini 3.1 Flash Lite as the preferred fallback model, and uses 0 ms delay unless the draft saves a different value.

1. Pick a strong primary model

   Fallback is not a license to choose a weak primary. Start with the model that best follows your instructions.

2. Enable fallback for latency, not decoration

   Use fallback when model response delay is a real caller problem.

3. Set the delay

   A 0 ms delay can start fallback immediately where fallback is active. Increase the delay when you want the primary model to have a chance to answer before fallback races it.

4. Choose the fallback model

   When fallback is enabled and nothing is selected, DialNexa tries to pick Gemini 3.1 Flash Lite as the default fallback where that model is available.

5. Check call evidence

   Review whether fallback won, whether the response was correct, and whether the user actually felt less delay.

Post-Call Extraction Model

The model used during a live call is not always the same model used for summaries and post-call extraction. DialNexa can run completed-call extraction through a text LLM path even when the live agent uses Speech to Speech.

Scenario	Behavior
Post-call extraction model selected	DialNexa uses the configured extraction model for post-call fields.
No extraction model selected	DialNexa falls back to the configured text LLM path for follow-up tasks.
Speech to Speech live agent	Post-call extraction switches to a text extraction model instead of trying to use the realtime speech provider for structured extraction.
Long or nuanced transcripts	Post-call extraction can use more reasoning budget than live turn generation because it runs after the call.

For field setup, see Post-Call Analysis.

Current Time Awareness

Use the current-time timezone control when the agent discusses dates, callbacks, business hours, deadlines, or relative time phrases. The selected timezone is saved on the agent and is included in runtime time context so the model can interpret phrases such as “today,” “tomorrow,” “in 2 hours,” or “after 5 PM” consistently.

For workflow follow-up scheduling, Time nodes still use their own resolved timezone path when parsing DateTime post-call analysis fields. See Workflow Time Nodes for callback scheduling behavior.

Predictive Preprocessing

Predictive preprocessing can pre-generate likely replies between turns for non-flow agents. The toggle is not shown for Conversational Flow Agents because flow behavior is explicit node logic.

Good fit	Poor fit
Repeated scripts, reminders, confirmations, and predictable objection paths.	Calls where the next line depends on a custom function result.
Agents with stable prompts and low variation between calls.	Agents with many dynamic variables in almost every sentence.
Short replies that commonly repeat.	Long exploratory conversations.

Prompt Caching

Prompt Caching can reuse a stable OpenAI prompt prefix across live conversation turns to reduce model latency and cost when the same instructions repeat. It is off by default and appears only for workspaces that DialNexa has allowlisted for the staged rollout.

Good fit	Poor fit
Large, stable prompts used across many similar calls.	Prompts that change often between calls or versions.
Campaigns where the same instructions repeat and only caller data changes.	Custom OpenAI-compatible providers that may reject OpenAI-specific cache fields.
OpenAI conversation models where cached-token evidence can be reviewed.	Speech to Speech paths or non-OpenAI live model paths.

DialNexa only sends prompt cache keys to the real OpenAI provider. The key is tied to the agent and version, then sharded by call so busy campaigns do not overload one cache lane. A cache miss only means the request runs normally. It does not change the response content.

When safe, DialNexa keeps the static prompt prefix stable and sends call-specific values in a separate system message. This helps campaigns reuse the same OpenAI prefix even when each call has different recipient values. The restructure is skipped and the prompt runs normally when any of these limits apply:

• The live model path is not OpenAI.
• An active fallback would send the same messages to a non-OpenAI provider.
• The agent is a Conversational Flow Agent.
• The prompt uses unsupported placeholder patterns such as escaped braces, positional placeholders, format specifications, nested fields, or malformed braces.
• A referenced dynamic value is missing, blank, or contains brace characters.
• A call-specific value would collide with runtime system markers such as time or language context.

Skipping the restructure is not an error. It protects prompt meaning and provider compatibility, but that call may not get the same cache benefit.

Empty Model Response Recovery

For normal cascaded turns, DialNexa treats whitespace-only, punctuation-only, and emoji-only model output as an empty response. It retries a clean empty response once by default without replaying the same response-cache entry. If the retry is also empty, the agent speaks a short request to repeat in the configured language where a localized message is available instead of leaving the caller in silence.

This recovery does not hide a real provider error, and it does not retry after meaningful speech has already started. Function calls also count as meaningful output. Review the recording and latency evidence when the caller hears a repeat request, because the extra model attempt can make that turn slower.

Model Problems And First Fixes

The agent does not call functions correctly

Lower temperature, then improve function descriptions, required fields, examples, and error handling. Change model only after the function schema is clear.

The agent gives shallow answers

Check whether knowledge content was retrieved and whether the prompt asks for the right depth. Then compare OpenAI and Google on the same call script.

The agent is slow

Confirm whether delay comes from transcription, model generation, custom functions, text to speech, or telephony. Use fallback LLM only when the model is actually the slow part.

The agent changes behavior between calls

Reduce temperature, tighten instructions, and remove conflicting prompt sections. Then compare models with the same test script.

The agent asks the caller to repeat after a pause

The model may have returned no meaningful text and exhausted its automatic retry. Review the transcript, model provider, fallback settings, and turn latency before changing the prompt.

How Model Behavior Affects Integrations

When an agent calls a function or prepares data for a workflow, the model is responsible for deciding when the action is appropriate and which values are safe to pass.

User goal	Model responsibility	Read next
Book or reschedule something.	Ask for missing fields before calling the booking action.	Functions, Google Calendar.
Update a CRM.	Separate confirmed caller facts from guesses.	HubSpot, Salesforce.
Send an email or WhatsApp message.	Avoid promising a message before the required recipient and content are known.	Email with Resend, WhatsApp with Wati.
Escalate a support case.	Summarize the issue, priority, and promised next step without inventing details.	Zendesk, Intercom, Slack.

Recap

Choose a language model after defining the call objective and required behavior. Test prompt adherence, latency, fallback behavior, functions, and difficult caller turns before changing live traffic.

Related Reading

Prompts And Welcome Messages

Write the instructions the model follows.

Functions

Give the model safe actions.

Custom Functions

Connect actions to your APIs.

Provider Selection Guide

Choose model, voice, and transcriber together.