Cohere
AIConnect to Cohere's enterprise language AI: chat and reasoning with the Command family, embeddings, best-in-class reranking for retrieval, classification, and dataset management.
What This Integration Enables
Cohere is tuned for retrieval-augmented workflows. Embeddings turn your content into vectors, reranking reorders retrieved candidates by real relevance rather than keyword overlap, and the Command models generate grounded answers over the top results. Classification and dataset actions round out the toolkit for building and maintaining a knowledge system. Reranking is the quiet differentiator: it is the step that turns a mediocre RAG answer into a good one by making sure the model reasons over the right passages. But retrieval and generation are only trustworthy when a person owns the answers that carry consequences. An orchestration layer decides which grounded answers return directly and which route to a human, and FlowRunner is built for that layer.
Without FlowRunner
With FlowRunner
Use Case Scenarios
Grounded Internal Q&A
Employees ask policy and process questions in a help channel. The agent embeds each question, retrieves candidate passages, reranks them so the most relevant rise to the top, and asks Command for an answer grounded in those passages with citations. Confident answers return directly with their sources. Questions touching sensitive policy are routed to the owning team, so self-serve never guesses on something that matters.
Support Ticket Classification
Incoming tickets need a category and a priority. The agent classifies each ticket with Cohere and routes it accordingly. The classification that used to require a person reading and sorting every ticket happens as they arrive, and the receiving team sees a triaged queue.
Search Relevance Upgrade
An existing search returns technically-matching but unhelpful results. The agent inserts a rerank step: it takes the top candidates from the existing search and reorders them by true relevance before showing them. Users find the right document faster, and when reranking surfaces nothing confidently relevant, the flow offers to route the question to a person rather than returning noise.
Human-in-Loop Highlight
A grounded answer is still an answer a person may need to own. When a Cohere-powered reply touches policy, commits the company to something, or comes back with weak retrieval support, FlowRunner routes it through a [human-in-the-loop](/concepts/human-in-the-loop/) step instead of returning it automatically. The agent pauses, assembles the question, the retrieved sources, and the draft answer, and sends it to the owning team via Slack. They confirm or correct. Retrieval and reranking do the heavy lifting; a person owns the answers that carry weight.
Agent Capabilities
24 actionsChat
3- Chat Generates a text response to a single prompt using a Cohere Command model via the v2 Chat API.
- Chat (Advanced) Sends a fully custom request to the Cohere v2 Chat API with a complete messages array (multi-turn conversations, multimodal image content for vision models), tool/function calling passthrough with strict tools, grounded documents, structured outputs via a response format object with JSON schema, citation options, safety modes, reasoning ('thinking') configuration, penalties and full sampling controls.
- Chat with Documents Generates a grounded answer to a question using the provided documents as context (Retrieval-Augmented Generation) via the Cohere v2 Chat API, Cohere's signature RAG capability.
Embeddings
1- Create Embeddings Generates embeddings for texts and/or images using the Cohere v2 Embed API. Supports Embed v4.
Rerank
1- Rerank Documents Ranks a list of documents by semantic relevance to a query using the Cohere v2 Rerank API, ideal for improving search quality and RAG retrieval.
Classification
1- Classify Text Classifies up to 96 texts using the Cohere Classify API, either few-shot with inline labeled examples (at least 2 examples per label recommended) or with a fine-tuned classification model that needs no examples.
Tokenization
2- Tokenize Text Splits a text (1 to 65,536 characters) into tokens using the tokenizer of the specified Cohere model.
- Detokenize Text Converts a list of token IDs back into the original text using the tokenizer of the specified Cohere model, the inverse of Tokenize Text.
Audio
1- Transcribe Audio Transcribes an audio file into text using Cohere Transcribe (state-of-the-art speech recognition across 14 languages including English, German, French, Spanish, Chinese, Arabic, Japanese and Korean) via the v2 Audio Transcriptions API.
Batches
4- Create Batch Creates and starts an asynchronous batch job that processes a previously uploaded and validated dataset of requests (dataset type 'batch-chat-v2-input', 'batch-embed-v2-input', 'batch-openai-chat-input' or 'batch-chat-input') with the chosen model.
- List Batches Lists the batch jobs of the authenticated user with their statuses, models, record counts and input/output dataset IDs.
- Get Batch Retrieves the current state of a batch job by its ID, including status (queued, in progress, completed, failed, canceled), record counts, token totals and the output_dataset_id for downloading results once completed.
- Cancel Batch Cancels an in-progress batch job by its ID. The batch moves to the canceling and then canceled status; records already processed are billed.
Datasets
5- Create Dataset Uploads a file to Cohere as a dataset for use with the Batch API or Embed Jobs API. Provide either the URL of an existing file (FlowRunner file URL or any public URL) or raw content as text (JSONL, CSV or TXT depending on the dataset type).
- List Datasets Lists datasets in the Cohere account with their types, validation statuses and download URLs.
- Get Dataset Retrieves a dataset by its ID, including its type, validation status, any validation error/warnings, and the dataset parts with download URLs for retrieving the data (e.g. batch job results or embed job outputs).
- Delete Dataset Permanently deletes a dataset from the Cohere account by its ID. Datasets are also automatically deleted 30 days after creation.
- Get Dataset Usage Retrieves the total dataset storage usage for the organization in bytes. Each organization can store up to 10GB of datasets across all users.
Embed Jobs
4- Create Embed Job Starts an asynchronous embed job that embeds an entire validated dataset of type 'embed-input', the way to embed large corpora beyond the 96-text limit of Create Embeddings.
- List Embed Jobs Lists all embed jobs of the authenticated user with their statuses, models and input/output dataset IDs.
- Get Embed Job Retrieves the details of an embed job by its ID, including status, model, input dataset and, once complete, the output_dataset_id containing the generated embeddings (downloadable via Get Dataset).
- Cancel Embed Job Cancels an active embed job by its ID. The embedding process is terminated; embeddings processed before cancellation are billed but partial results are not made available.
Models
2- List Models Lists the models available through the Cohere API with their compatible endpoints, context lengths, deprecation status and supported features.
- Get Model Retrieves the details of a specific Cohere model by its name, including compatible endpoints, context length, deprecation status, tokenizer URL, supported features and default sampling parameters.
Start building with Cohere
$100 in credits. No card required. Connect in minutes.