SambaNova

Run chat completions, single-prompt generation, and embeddings on SambaNova Cloud's RDU hardware through an OpenAI-compatible API. Agents call Llama, DeepSeek, and Qwen models with tool calling and structured JSON output, and transcribe audio with Whisper.

5 actions API key available

SambaNova website ↗ Platform Documentation ↗ Capability data verified 2026-07-25

A finished sales call recording lands in storage from the meeting platform

Transcribe Audio converts the recording, with a language hint and a prompt seeded with product names and acronyms

Simple Chat in JSON mode extracts commitments, pricing mentions, objections, and next steps from the transcript

Create Embeddings indexes the transcript chunks into the deal knowledge base

The agent flags every extracted commitment that involves a price, a discount, or a date

The account owner verifies flagged commitments against the recording before anything writes to the deal record

Verified fields update the CRM and the call summary posts to the deal channel

What This Integration Enables

SambaNova Cloud runs open models, Llama, DeepSeek, Qwen, on RDU hardware behind an OpenAI-compatible API, and it packs the full ear-to-action loop into one connector: Whisper-Large-v3 transcription in, chat and reasoning models to extract meaning, embeddings to make it retrievable. That last-mile combination is what FlowRunner exploits. Most operational knowledge in a company is spoken, on sales calls, support lines, and site walkthroughs, and it dies there because nothing structured ever hears it. - Transcribe Audio handles 50+ languages, with language hints and prompts that teach it your product names and acronyms - Simple Chat turns transcripts into summaries, classifications, and JSON-mode structured fields in one step - Chat Completion adds multi-turn messages, multimodal content parts, tool calling, and structured response formats - Create Embeddings feeds semantic search, RAG, and deduplication pipelines, vectors returned in input order - List Models keeps model selection current with the account's available chat, embedding, and speech models - Spoken commitments become recorded facts only after [human-in-the-loop](/concepts/human-in-the-loop) verification

Without FlowRunner

Calls evaporate What was said on the call survives as three bullet points typed from memory two days later

Voice is a dead medium for automation Recordings pile up in storage because nothing downstream can read them

Transcripts get treated as truth A misheard number from speech-to-text flows into a record, and now the system is confidently wrong

With FlowRunner

Every call becomes structured data Transcription, extraction, and indexing run as one pipeline the moment the recording lands

Workflows get ears Voice notes, meetings, and support calls feed the same flows that documents and emails already do

The record checks the recording Extracted commitments route to a person who confirms them against the audio before they become facts

Use Case Scenarios

Sales calls that file their own paperwork

Call recordings flow in after every meeting. The agent transcribes each with Transcribe Audio, extracts structured fields with Simple Chat in JSON mode, deal stage signals, objections, competitor mentions, promised follow-ups, and posts the summary to the deal channel in [Slack](/integrations/slack). Routine fields update [HubSpot](/integrations/hubspot) directly; anything involving money or dates waits for the rep's confirmation. Pipeline reviews start reading what was actually said instead of what got remembered.

Field voice notes that become work orders

Technicians on site record thirty-second voice notes instead of typing on a ladder. The agent transcribes each note, extracts site, issue, severity, and parts needed into JSON, and creates the task in [ClickUp](/integrations/clickup) with the original audio linked. Notes the model could not confidently parse route to the dispatcher with the transcript and audio side by side. The field team talks; the system types; the dispatcher only touches the ambiguous ones.

A support call archive you can actually ask questions of

Historic support calls sit as audio nobody will ever replay. The agent works through the backlog: Transcribe Audio per call, Create Embeddings per transcript chunk, vectors into [Pinecone](/integrations/pinecone) with call metadata attached. From then on, "have customers complained about the new billing flow on calls, not just tickets" is a retrieval query answered with cited transcript passages, not a guess. The multilingual model means the Spanish and German call archives join the same index.

Human-in-Loop Highlight

Transcribe Audio is the point where this connector manufactures facts, and speech-to-text fails in the worst possible way: fluently. Whisper does not stumble when it mishears; it produces a clean, confident sentence in which "fifteen" became "fifty" and the discount you offered doubled. When the transcript feeds an extraction step and the extraction feeds a CRM field or a follow-up email, a mishearing becomes a commitment your company appears to have made. So the pipeline never promotes money or dates from audio to record without a person. The flow presents the claim with its evidence: "Extracted from today's Meridian call: 'agreed to 15% discount on annual, decision by March 3.' Transcript excerpt and audio timestamp attached. Confirm before CRM update?" The rep clicks the timestamp, hears their own voice, confirms or corrects. Transcription runs at machine speed across every call the company makes; what the company officially said is confirmed by someone who was in the room.

Agent processes routinely

Detects exception requiring judgment

Clear match Continues automatically

Ambiguous Routes to human via email

Human decides

Agent resumes with decision

Agent Capabilities

5 actions

Chat

Chat Completion Generates a chat completion with a full OpenAI-compatible messages array: multi-turn conversations, multimodal content parts, the Llama, DeepSeek, and Qwen families, sampling controls, structured outputs via a response format object, and tool and function calling. Returns the raw response with choices, tool calls, and token usage.
Simple Chat Sends a single prompt and returns just the reply text, with an optional system prompt, sampling controls, and JSON mode for structured output. The workhorse for summarizing, classifying, drafting, and extraction steps.

Embeddings

Create Embeddings Converts one or more texts into embedding vectors with a SambaNova embedding model such as E5-Mistral-7B-Instruct, for semantic search, RAG, clustering, deduplication, and similarity scoring. Vectors return in input order with token usage.

Audio

Transcribe Audio Transcribes an audio file with Whisper-Large-v3 across 50+ languages. Audio is pulled from a FlowRunner file URL or any public URL, up to 25 MB, with an optional language hint and a prompt that improves accuracy on names, acronyms, and uncommon words. The fact-manufacturing step this page's human gate verifies.

Models

List Models Lists the models available in the account, chat and reasoning, embedding, and Whisper speech-to-text, with owners and context window sizes. The discovery step before committing a flow to a model ID.

Frequently Asked Questions

What can FlowRunner do with SambaNova?

FlowRunner agents can run Chat Completion, Simple Chat, and Create Embeddings in SambaNova, plus 2 more actions.

Does connecting SambaNova to FlowRunner require OAuth?

No. SambaNova connects to FlowRunner with an API key, no OAuth flow required.

Can SambaNova trigger a FlowRunner workflow automatically?

SambaNova doesn't currently expose triggers in FlowRunner. It connects as an action step inside workflows started by another trigger.

Start building with SambaNova

$100 in credits. No card required. Connect in minutes.

Start Building Free Talk to us