Troubleshooting by symptom

Check in this order:

1. gpt-5 model with tools. If the agent uses tools (tool calls or a flow) with a gpt-5 model, set Reasoning effort to NONE in the Generation section. Without an explicit effort, the call fails and the bot stays silent. See Generation.
2. Pending human validation. If the flow contains a Human validation block, the conversation waits for an operator's decision. In the tester, the message Waiting for supervisor approval… is shown. See Human approvals.
3. Full conversation. In Strict memory, the conversation stops when the window is full. The tester then shows the token fill and New conversation. Switch to Rolling memory or enlarge the Conversation window. See Memory mode.
4. Linked flow not found. If the Linked flow section shows Orphan reference, the agent cannot run its flow. Click Detach, then choose an existing flow. See Orphan reference.
5. Error in the trace. Open the trace of the exchange: the failed step shows its error message.

• The agent is not saved. The tester and calls use the saved configuration. Click Save and wait for Saved.
• The flow is not saved. The auto-saved mention means your changes are only kept in your browser. Click Save in the flow editor to create a new version. See Save and publish.
• The flow version is pinned. If the agent runs a specific version (v8, for example), a new flow version does not apply. In the agent's Linked flow section, pin the new version or choose Latest — follows the flow's changes, then save the agent.
• The test conversation is old. Click Reset in the tester to start a fresh conversation.

• Vague prompt. Structure the system prompt into role, goal, behavior, tone and guardrails, and say what to do when the agent does not know ("offer a transfer rather than making something up"). See Write an agent's prompt.
• Temperature too high. For support or an FAQ, a value between 0.1 and 0.3 gives more consistent answers.
• Variable not provided. A {{…}} variable in the prompt received no value. In debug mode, the transcript flags a variable used in a prompt but missing when entering the node. Simulate values with the tester's Test variables.
• Too many linked tools. Each linked tool is offered to the model at every turn: a short catalog makes its choices more reliable.

1. In Knowledge, check that the document has the Indexed badge, not Processing or Failed. See Track your documents.
2. Ask the question in Test semantic search…, with a customer's words. If "No segment found." is shown, rephrase the document or add to the base. See Test the search.
3. In the agent, Knowledge (RAG) section, check that Enable RAG is on and that the right Knowledge base is selected.
4. Lower the Min score if too few passages come up; set Search to HYBRID if the question uses different words from the document.
5. Say in the prompt when to consult the base.

After changing the Embeddings configuration, run Re-index in the base Properties: without re-indexing, older documents are no longer found correctly.

In a voice call, RAG does not apply: the agent consults the base only through the KB search — base name tool. Open the agent's Knowledge (RAG) section: the message "Base also available on voice calls, through the tool … connected to the agent." confirms that the tool is connected. Otherwise, save the agent so that Flowtative creates and connects it. See Connect a base to an agent.

The search tool keeps the settings it had when it was created. Edit it in Settings > Tools. See The search tool added automatically.

Raise the Min score, reduce Max results, or split the base by scope (a "Customer service FAQ" base, a "Product catalog" base): one targeted base per agent gives more precise answers.

This is typical of dictation (contract number, address). In Turn-taking, try the Detection mode Smart Turn (AI judges each pause) or increase the Resume delay (VAD). With Deepgram, Endpointing also sets the silence after which a sentence is considered finished. For digit entry, offer the keypad instead: set the Expected length in the block's Phone keypad (DTMF). See Turn-taking and Phone keypad.

Increase Min. words to interrupt in Turn-taking: the caller will need to say more words to interrupt the bot. In Realtime mode, interruption is set with Interruption (barge-in). For keypad presses, turn off The 1st key interrupts the bot in the relevant block.

In TTS — voice synthesis, set the Speed (from × 0.70 to × 1.20), keep an understated Style and check the rendering with Listen with these settings on a real sentence from your journey. See Configure a callbot's voice.

Add an entry in Text substitutions: SAV → S A V, RDV → rendez-vous. For a more general pattern, use a regex rule. See Text substitutions.

• With Deepgram, add these words to Keywords (boost), with a weight if needed (ViaDialog:2).
• With VIASPEECH, add them to Hotwords (vocabulary).
• Check that the Language matches your callers' language (fr).

Then compare what the caller said (replay of the Caller track) with what was transcribed, in Transcripts.

Open the trace of a slow exchange or the agent's Latency breakdown (average) in Observability to see which step takes the time:

See Agent panel.

In End of call, keep Wait for the end of the message before hanging up on. See End of call.

Add this tool's ID to End-of-conversation tools (End of call section): calling it will mark the end of the conversation.

See Test an agent.

In Settings > Telephony & Trunk, Inbound numbers part:

1. The number is in the list and an agent is connected to it (the link line is lit).
2. The line switch is on: a disabled line appears dimmed.
3. The connected agent is the right one. Otherwise, click its name and choose another.

From the agent editor, the Inbound number section also lists the numbers that route calls to it, with the disabled mention where relevant. See Telephony.

Check the SIP trunk status. Drift: open the trunk with Edit, then save to realign the platform. Missing from the platform: the trunk no longer exists, delete it and create it again. Not verified: try again later with Refresh. Also check that the called number is in the trunk's Routed numbers. See Synchronization statuses.

The number is connected to another agent. In Inbound numbers, click the agent's name and choose the right one. The next calls reach the new agent.

The transfer is set on the Hang up block, in End parameters: the Call transfer shortcut adds the Transfer number and the Transfer mode. Check that the flow actually reaches this block (diagnostic log), that the number is complete and that the mode fits: BRIDGE keeps the bot on the line, REFER takes it out of the call. Transfer only applies to voice calls. See Transfer to an advisor.

See Track and manage a campaign and The calling window.

• Excluded · invalid number: the record has no valid number. Fix it in the contact base, then reset the contact.
• Prospecting: records without positive consent in the consent column (empty, "non", "no", "false", "0") are excluded.
• Legal cap: contacts already solicited 4 times in 30 days are excluded from the audience.
• Done after a Voicemail: with Voicemail closes the contact checked (default), a voicemail ends the contact.
• No attempts left: all attempts have been used.

The Contact details panel shows, under Next call, why a contact will no longer be called. See Contact details.

A campaign moves to Done when you close it or when its end date has passed. It no longer calls. To start again with the same settings, click Duplicate. See Campaign lifecycle.

This is expected: as long as it is not closed and its end date has not passed, scheduled retries can still go out. Click Close if you want to stop it for good.

A retry never falls outside the calling window: if the delay takes it out of the current range, it is moved to the next range. The contact's Next call shows the planned date. See Retries.

The agent writes the business result only if it has a tool that updates the base and if its prompt asks it to. Add a column with an explicit name (resultat, statut, issue…) and ask the agent to write the conversation outcome in it. See The outcome on each record.

1. In the agent's Capabilities section, Tool calls is on.
2. The tool appears in Linked tools with the Linked status. To create means no tool has this ID; Error means it could not be loaded (click Retry, do not recreate it).
3. The tool's Description says, from the AI's point of view, when to use it, and each parameter is described with a format example.
4. In a flow, the function's When to call it? description is explicit.

See Link tools to an agent.

A function runs a tool only if its name exactly matches the tool's ID, case included. Otherwise, it only serves as a transition, with no visible error. If you changed the tool's ID after using it in a flow, fix the function name. See Use a tool in a flow.

Test it in Settings > Tools with Test: the result shows Success or Failure and the response received. Check the URL, the headers and the ${name} parameters. In Observability, the agent's Tool reliability gives the success rate of each tool, and the trace shows the beginning of the response returned. An MCP tool marked missing no longer exists on its server.

These blocks run a tool as soon as they are entered, without going through the AI. If the required tool does not exist in your workspace, the block panel shows a warning and the flow fails on load. For a Datastore block, click Create the tool now. In the agent, a banner also flags a tool called by the linked flow but not found. See Flow tools added automatically.

This is expected: their settings are kept with the flow, but are not executed during a conversation. To call an API, create an HTTP / API tool and call it from a function or an MCP tool block. To route a journey, use a function with a Conditional Output. See HTTP request.

The Update operation replaces the whole record: a field left empty is lost. Fill in every field, for example from the result of an earlier search. See Datastore.

1. The webhook is active in Settings > Webhooks.
2. It is checked in the agent's Webhooks section, and the agent has been saved.
3. It does not show "Does not listen to BOT_SESSION_ENDED — it will receive nothing".
4. The agent is a callbot: today, only callbot call sessions are sent.
5. The Failed deliveries log contains no error.

See Webhooks.

• Automation active is on, or the list switch is on.
• The Upcoming runs shown match the expected time zone and time.
• The agent's flow is Autonomous: only blocks with no customer exchange can be reached from its entry. See Make a flow autonomous.
• The automation's History shows Failed (with the error message) or Rejected, for example for an API call on an automation that does not have the API call trigger.

See Automations.

The agent is set up for chat. Click Enable voice (callbot) to add the group. The Call recording option and end-of-session webhooks also only concern callbots. See Configure a callbot's voice.

Your role does not allow this action. For example, resetting a campaign contact is reserved for the Administrator and Service admin roles, and changing another member's role can be refused ("Could not update the user (403)."). Ask an administrator to check your role in Team. See Roles.

This is by design: your role is shown read-only on your row. Ask another administrator.

A platform skill (platform badge) is read-only for your company, as is a skill published from a wiki. To adapt a platform skill, create a company skill with the same name: that is the one your agents will load. See Skills.

The version button (v12, for example) appears after the flow is saved for the first time. A flow that was never saved has neither history nor a version to link to an agent.

Real-time supervision is not yet enabled on your workspace. Contact your ViaDialog representative. Finished conversations remain available in Transcripts.

The Call recording option must be on in the agent's Capabilities, the agent must be a callbot, and recording must also be enabled on the platform side. Without a recording, the page shows "No recording available for this conversation."

Its Audit log is not on: only its volume is known. Turn it on in the agent's Capabilities section. See Observability.