Integration Architecture Overview
Choose the right integration pattern—Pre-Call Fetch, Mid-Call HTTP Tools, MCP servers, Knowledge Bases, Webhook Nodes, or `n8n.talkar.in`—for every CRM and backend use case.
Integration Architecture Overview
As a Talkar builder, choosing where an API call or lookup happens matters just as much as writing the payload. Voice conversations are latency-sensitive: an avoidable 2-second blocking HTTP call mid-sentence feels broken to a caller, whereas loading the wrong data upfront bloats your prompt.
Use this guide to pick the right integration mechanism before wiring up HubSpot, Zoho, Salesforce, custom ERPs, or internal middleware.
Six Integration Patterns at a Glance
| Pattern | When It Runs | Latency Impact | Best For |
|---|---|---|---|
| Pre-Call Data Fetch | Before the Start Node speaks | Zero mid-call delay (adds tiny setup time before greeting) | Loading caller name, account status, open order ID, or language preference by phone number before "Hello" |
Mid-Call HTTP Tool (http_api) | While the agent is talking, when the LLM calls a function | Caller hears silence or your configured filler message until response returns (default timeout 5s) | Checking live appointment slots, looking up a tracking number provided mid-call, creating a support ticket |
| Knowledge Base (Chunked RAG / Full Doc) | Indexed upfront; searched mid-call via retrieve_from_knowledge_base (returns chunks or full text on demand) | Fast tool-call lookup when invoked | FAQs, pricing sheets, return policies, troubleshooting manuals, warranty terms |
MCP Server (mcp) | Session connected at call start; tools called dynamically by LLM | Depends on remote MCP server (timeout_secs 30s / SSE 300s defaults) | Multi-tool external ecosystems (CRMs, calendars, ticket queues) exposed over Streamable HTTP |
Post-Call Webhook Node (webhook) | Asynchronously after the call ends (after QA & Post-Call Intelligence finish) | Zero impact on call audio | Firing guaranteed CRM status updates, sending extracted_data, QA annotations, recordings & transcripts with retries (up to 5 attempts) |
| Talkar Service & Automation APIs | Triggered externally over REST or consumed after calls | Zero impact on call audio | Launching outbound calls, streaming live SSE transcripts, embedding widgets, running bulk campaigns, wallet/account automation |
End-to-End CRM Data Flow
Recommended Pattern: n8n.talkar.in as Middleware
For multi-step CRM workflows (such as searching HubSpot by phone, falling back to creating a contact, formatting dates, and logging to Slack/WhatsApp), avoid chaining multiple direct API tools inside the voice LLM loop.
Talkar operates an internal workflow automation instance at https://n8n.talkar.in:
- Mid-Call Actions (
http_api->n8n.talkar.in): Expose one fast Webhook trigger in n8n that performs lookup/transformation behind the scenes and returns clean, concise JSON back to the voice agent within the5000 msvoice timeout. - Dynamic Call Transfers (
transfer_callresolver ->n8n.talkar.in): Point a dynamic transfer resolver webhook to n8n so your routing logic can check business hours, agent availability, or CRM account ownership and return{"transfer_context": {"destination": "<number_or_sip>", "custom_message": "..."}}. - Post-Call Fan-Out: Receive post-call transcripts, recordings, and extracted variables in n8n to update Zoho/Salesforce/Google Sheets asynchronously without risking live call latency.
Protect secrets in the Credentials Vault
Never hardcode CRM bearer tokens or API keys inside headers JSON when you can store them in the Credentials Vault (/api/v1/credentials). Vault secrets are write-only via API, masked in tool test previews, and hidden from LLM tool outputs.
Decision Matrix for Client Builds
- "I need the bot to greet the caller by name and know their unpaid invoice amount." -> Use Pre-Call Data Fetch on the Start Call node (or pass values directly in
initial_contextvia External Trigger API / Campaign CSV). - "While talking, the caller gives an order ID and asks where their package is." -> Use a Mid-Call HTTP Tool (
http_api) with`{{order_id}}`path or body parameters and a shortcustomMessagelike"Let me check that order for you.". - "The bot needs to answer questions from a 20-page PDF policy handbook." -> Upload the document to the Knowledge Base and attach it to the relevant Agent nodes.
- "After every call finishes, I need a guaranteed POST to our CRM with transcript, recording URL, QA tags, and extracted fields." -> Add a Post-Call Webhook Node (
webhook) on the canvas (no edges required) so Talkar automatically delivers the payload with built-in retries after Post-Call Intelligence finishes. - "Our client's backend software wants to trigger outbound calls programmatically and embed the agent on a web portal." -> Use the Talkar Service API Reference (Trigger API, Live SSE transcripts, Embed Tokens, and Automation API).
Turns, Silero VAD & Interruptions
Plain-English UI guide for Voice Sensitivity presets, Start/Stop turn strategies, Silero VAD vs. Provider VAD, the 3 user muting layers, Smart Turn, and Realtime S2S turn ownership.
Pre-Call Data Fetch & Dynamic Variables
Load CRM records into `initial_context` before the first greeting using Pre-Call Fetch, Campaign CSV columns, or Trigger API payloads.