AI21 Labs
AICall AI21's Jamba models from your flows. Agents run chat completions with tool calling and JSON mode, ask one-shot questions, and answer questions grounded in supplied documents using Jamba's 256K token context window.
What This Integration Enables
The standard architecture for asking questions about documents assumes a pipeline: chunk, embed, store, retrieve, then answer. AI21's Jamba models remove the middle for a large class of work. With a 256K token context window and a native documents parameter, Ask With Documents takes entire contracts, meeting transcripts, policy manuals, or scraped pages inline and answers from that content instead of the model's general knowledge. For FlowRunner agents, that means document intelligence with fewer moving parts to govern and a clear provenance story: the answer came from these documents, supplied in this run. Retrieval infrastructure still earns its place when the corpus runs to millions of documents; the point is that the large middle of document work, one contract or one transcript at a time, no longer requires it. - Answer questions over long source material without standing up a vector database - Ground responses in supplied documents with per-document metadata the model can cite - Extract structured fields from contracts and reports with JSON mode - Summarize, classify, draft, and rewrite with one-shot Ask Jamba calls - Run multi-turn assistants with tool calling through Chat Completion - Use Jamba as the answering step behind any search or scrape that produces top results
Without FlowRunner
With FlowRunner
Use Case Scenarios
Research questions answered from the pages, not about them
An analyst drops a question and a list of URLs into the intake form. The flow scrapes the pages with [Firecrawl](/integrations/firecrawl), passes the full content to Ask With Documents with each page's URL and date as metadata, and posts the grounded answer to [Slack](/integrations/slack) with the source list attached. The model is instructed to answer strictly from the supplied pages, so a claim that is not in them comes back as absent rather than improvised. Answers log with their document list, and a challenged conclusion traces back to the pages it stood on.
The contract intake line
When a signed agreement arrives in [Google Drive](/integrations/google-drive-api), the flow reads the content and runs Ask With Documents with JSON mode: renewal date, notice period, liability cap, payment terms. The structured record writes to [Airtable](/integrations/airtable) with the quoted passages beside each field, and the notice deadline lands on the contract calendar. The whole contract fits in context, so the clause on page 41 gets the same attention as the one on page 2.
Support replies drafted from the actual policy
A ticket arrives and Ask Jamba drafts the reply, with the current policy document supplied as grounding so the draft reflects what the policy says today, not what the model remembers policies usually saying. Routine drafts queue for a one-click send through [Gmail](/integrations/gmail-service), while anything touching refunds or account closure holds for a person. Each interaction logs to [Google Sheets](/integrations/google-sheets) with the model, tokens, and disposition.
Human-in-Loop Highlight
The dangerous output here is a date. When Ask With Documents reads a contract and reports that the notice period is 60 days before the renewal on March 1, that answer becomes a calendar entry, and the calendar entry becomes the day the company acts or fails to act. Miss the real window because the model conflated the initial term with a renewal term, and the contract renews itself for a year at the old price; there is no undo for a lapsed notice deadline. So the FlowRunner intake flow treats extraction as a draft, not a fact. Every extracted term posts with the exact passage it came from, side by side, and legal ops confirms the reading before the deadline is booked and the tracker updated. Confirmations take a minute because the passage is right there; the model did the reading of the other forty pages. What never happens is a contractual deadline entering the calendar on the model's word alone.
Agent Capabilities
3 actionsConversation
1- Chat Completion Sends a full chat completion request to a Jamba model, a hybrid Mamba-Transformer architecture with a 256K token context window, using a complete messages array. Supports multi-turn conversations, system instructions, tool calling, JSON mode, grounding documents, and multiple candidate completions, returning text, tool calls, finish reason, and usage. Note the API caps max_tokens at 4096, and n must be 1 when tools are supplied.
One-Shot Tasks
1- Ask Jamba Sends a single prompt and returns the assistant's answer as text, building the messages array from an optional system prompt and a user message. The fast path for summarization, classification, drafting, and extraction, with JSON mode for structured output.
Grounded Answers
1- Ask With Documents Answers a question grounded in documents you supply through Jamba's native documents parameter, so the response is based on that content rather than general knowledge. Entire contracts, transcripts, and knowledge-base articles pass inline within the 256K window, no vector database required, and each document can carry author, URL, and date metadata the model can cite. The answering step of a retrieval pipeline, and the action this page's human gate keeps honest.
Frequently Asked Questions
What can FlowRunner do with AI21 Labs?
FlowRunner agents can run Chat Completion, Ask Jamba, and Ask With Documents in AI21 Labs.
Does connecting AI21 Labs to FlowRunner require OAuth?
No. AI21 Labs connects to FlowRunner with an API key, no OAuth flow required.
Can AI21 Labs trigger a FlowRunner workflow automatically?
AI21 Labs doesn't currently expose triggers in FlowRunner. It connects as an action step inside workflows started by another trigger.
Start building with AI21 Labs
$100 in credits. No card required. Connect in minutes.