Introducing FileFeed AI Assistant: Set Up a Complete Data Pipeline From a Chat Conversation

The onboarding problem that every data team hits

Setting up a new data pipeline is never just one step. You create a client. You define a schema. You figure out how the source columns map to your target fields. You configure transformations for the columns that do not match cleanly. You set up webhooks so your system knows when new data arrives. Then you write the integration code: an endpoint to verify webhook signatures, a service to fetch processed data, a loop to paginate through records.

For an experienced engineer, this takes an afternoon. For a new team member, it takes a day or more of reading documentation, asking questions, and debugging configuration. For an operations lead who knows the business logic but is not comfortable writing JSON schemas by hand, it can block them entirely.

We kept hearing the same thing from customers: "I know what I want the pipeline to do. I just need someone to translate that into the right configuration." That someone is now an AI.

What the AI Assistant actually does

The FileFeed AI Assistant is embedded directly in the dashboard. Open it from any page, and you are in a conversation with an AI that has full access to your FileFeed workspace. It can read your existing clients, schemas, pipelines, and pipeline runs. It can create new resources. It can analyze files you upload. And it does all of this through plain language conversation.

This is not a chatbot that answers FAQs. It is an agent that executes real operations on your account, with your approval at every write step.

File analysis built into the chat

You can upload CSV, JSON, or XML files directly in the chat. Drop a file and the assistant analyzes it: parses the structure, detects column types, counts rows, identifies data quality issues, and shows you a summary. Upload two files one representing what your client sends, one representing what your system expects and the assistant compares them side by side. It identifies matching columns, suggests field mappings, and recommends transformations where the data types do not align.

This file analysis is not a separate feature you navigate to. It happens inline, in the same conversation where you are planning the pipeline. You upload, the assistant analyzes, and you continue talking about what to build.

End-to-end pipeline creation from conversation

When you tell the assistant to set up a pipeline, it follows a structured workflow. It checks if the client already exists. It asks for the client name if you are creating a new one. It proposes a schema based on the target file structure, showing you the field definitions in a table so you can review before it creates anything. It builds the pipeline with field mappings from source to target, including any transformations the data needs. It asks if you want webhooks. At every step, it confirms with you before executing.

The assistant does not guess. If it is not sure which source column maps to which target field, it asks you. If the source file has columns that do not appear in the target, it tells you and asks how you want to handle them. If there is an existing pipeline that conflicts, it explains the conflict and offers options.

Personalized integration guide at the end

This is the part that surprised even us during development. After the assistant finishes creating all the resources, it does not just say "done" and leave you to figure out the next steps. It generates a complete, personalized integration guide tailored to the exact resources it just created.

The guide includes an architecture diagram showing the data flow from SFTP upload to your system. It includes the actual SFTP host and username for your new client. It includes a working webhook handler in Node.js with HMAC-SHA256 signature verification using the standard x-sftpsync-signature header. It includes TypeScript code using the @filefeed/sdk to fetch processed data with pagination and acknowledge completed runs. It includes a testing checklist specific to your pipeline.

This is not a generic template. The code references your actual pipeline ID, your schema name, your client name. A developer can copy the code, paste it into their project, add the API key, and it works.

A real setup session, start to finish

Here is what a real pipeline setup looks like using the AI Assistant. A customer onboarding manager receives two files from a new partner: the partner's data export and the internal format their system expects.

1. Open the FileFeed AI Assistant from the dashboard sidebar.
2. Upload both files and type: "The first file is what Samsung sends us, the second is what our system needs. Set up a complete SFTP pipeline."
3. The assistant analyzes both files and presents a summary: 5 source columns, 3 target columns, 3 suggested mappings, 2 unmapped source columns to ignore.
4. The assistant asks: "Which client is this for?" and shows existing clients plus "Create new client" as an option.
5. You type: "Create a new one called Samsung."
6. The assistant creates the client with hosted SFTP auto-provisioned. Confirms: "Samsung client created. SFTP host: sftp.sftpsync.io, username: samsung-demo."
7. The assistant proposes the schema based on the target file: id (number, required), site_name (string, required), site_description (string). You approve.
8. The assistant creates the pipeline with field mappings: ID → id, name → site_name, des → site_description. Two source columns (sourceField2_1, sourceField2_2) are excluded as discussed. You approve.
9. The assistant asks about webhooks. You say yes and provide a URL. It creates the webhook and reminds you to save the secret.
10. The assistant outputs a full integration guide with architecture diagram, SFTP credentials, webhook handler code, SDK data retrieval code, and a testing checklist.

Total time: about three minutes. Total dashboard forms filled out manually: zero. The entire setup happened in a conversation.

What it can do beyond pipeline setup

The setup wizard is the most dramatic example, but the assistant is useful for day-to-day operations too. It has access to your entire workspace and can answer questions that would otherwise require navigating through multiple dashboard pages.

• Investigate failed runs: "Show me all failed pipeline runs from the last week" returns a filtered list. "What went wrong with run X?" shows the error details. "Reprocess it" triggers a retry, with your approval.
• Search processed data: "Did the name John Smith appear in any processed file this month?" uses the search tool to scan across pipeline run data without downloading files.
• Understand pipeline configuration: "How is the Samsung pipeline configured? What fields are mapped?" retrieves the full pipeline details including mappings, transforms, and connected webhooks.
• Monitor volume and performance: "How many files did we process for ACME this week? Any failures?" gives you an instant operational snapshot.
• Modify existing resources: "Add a new field called 'department' to the Samsung schema" or "Change the mapping for column X to target field Y" configuration changes through conversation.

How it works under the hood

The AI Assistant runs on Anthropic's Claude model with a carefully designed agentic architecture. When you send a message, the assistant processes it through an agentic loop that can call tools, analyze results, and decide what to do next up to 15 iterations per message. It streams responses token by token, so you see the assistant's thinking in real time.

The tool system includes over 25 operations spanning every FileFeed resource. Read operations (listing clients, retrieving pipeline details, fetching processed data) execute immediately. Write operations (creating, updating, deleting) go through a human-in-the-loop approval gate. The assistant pauses, shows you what it wants to do, and waits for your decision. You can approve once, deny, or approve all future calls to that tool type for the session.

File attachments are uploaded to temporary storage, analyzed server-side using the same parsing engine that processes pipeline files (CSV via PapaParse, JSON, XML via xml2js), and made available to the AI through dedicated analysis tools. The AI never sees your raw file content in its prompt. Instead, it calls tools that return structured summaries: column names, inferred types, sample values, data quality issues. This keeps conversations efficient even with large files.

Security and access control

The assistant operates within your workspace scope. It can only see and modify resources that belong to your workspace, using the same JWT authentication as the rest of the dashboard. Sensitive fields like SFTP passwords, webhook secrets, and private keys are automatically redacted before they reach the AI model. The assistant never sees credentials and cannot leak them in a response.

File attachments uploaded through the chat are stored temporarily with a 24-hour expiration. They are scoped to the conversation and automatically cleaned up. No uploaded data persists beyond the session unless you explicitly create a pipeline and process files through it.

Who this is for

We built the AI Assistant for three personas that kept showing up in our customer conversations.

Operations leads who understand the business requirements which clients send data, what format it arrives in, what the target schema should look like but do not want to translate that knowledge into JSON schemas and mapping configurations by hand. The assistant lets them describe what they want in plain language and handles the technical translation.

Engineers during initial setup who know exactly what they need to configure but want to move faster than form-by-form dashboard navigation allows. Upload two files, describe the pipeline, approve the steps, and get working integration code. Ten minutes of dashboard work becomes three minutes of conversation.

Support and customer success teams who need to quickly investigate pipeline issues, check run statuses, and pull processed data for customers. Instead of clicking through multiple dashboard pages, they ask a question and get an answer. "Why did Samsung's last run fail?" is faster than navigating to Pipeline Runs → filter by client → click the run → read the error message.

Available now in every FileFeed workspace

The AI Assistant is live in the FileFeed dashboard for all workspaces. Open it from the chat icon on any page. There is nothing to install or configure. If you have an existing automated file feed running, the assistant already knows about your clients, schemas, and pipelines.

If you are new to FileFeed, the assistant is the fastest way to go from zero to a working pipeline. Read our introduction to FileFeed for the full product overview. Sign up for a free account, upload your files, and tell the assistant what you need. It will handle the rest.

For teams that want AI capabilities outside the dashboard too, the FileFeed MCP server brings the same tools into Claude Desktop, Cursor, and VS Code. Same API, different interface.