GPT-trainer

Build, train, and operate GPT-trainer chatbots and their multi-agent configuration from FlowRunner. Agents keep the knowledge base current as source files and pages change.

1 trigger 30 actions API key available

Platform Documentation ↗ Capability data verified 2026-08-01

A documentation page is published and the CMS fires its webhook

Agent matches the published URL against List Data Sources to find the training source holding the previous version

Agent calls Retrain URL Data Sources on that source UUID so GPT-trainer re-scrapes the page

Agent polls List Data Sources until the status leaves downloading, chunking and embedding and reaches success

Agent opens a session, asks three canary questions the page is supposed to answer, and reads the citations back with List Messages

Agent posts the old answer, the new answer and the cited source titles into the documentation channel

The documentation owner confirms the bot now answers correctly, or decides the topic needs a pinned Q&A source instead of a scraped page

What This Integration Enables

GPT-trainer earns its name. The connector's centre of gravity is the training corpus, and everything else is arranged around keeping it correct. Sources arrive as uploaded files, scraped URLs or hand written Q&A pairs, and each one reports a granular indexing status through assigned, await, downloading, chunking, embedding, running and success, with distinct error values that tell you whether you hit a storage limit or a token limit rather than leaving you to guess. Retrain URL Data Sources re-scrapes link sources and returns a per UUID outcome, so one bad id does not sink the batch. Source tags group the library by product line, language or department, and a User-Facing agent can be restricted to specific source UUIDs or opened to the whole library. Update Data Source Title exists for a reason worth noticing: titles are generated automatically and the title is what the chatbot shows when it cites that source, so renaming one is how a citation is made readable to a customer.

Underneath that sits a multi-agent chatbot with the same typed roles you would expect: User-Facing agents answer from the source library, Background agents monitor and tag without replying, Human Escalation routes to a person, Pre-Canned returns fixed replies, and Spam Defense filters abuse. New agents are created disabled, so Update Agent is the operation that puts one into service. And when a conversation needs to leave the platform, Get Session Transcript returns the whole exchange as plain alternating User and Assistant lines in a single call, which is a small convenience that removes a lot of post processing when the destination is a ticket, a summary or an email to a person who needs to decide something.

Without FlowRunner

Docs and bot drift apart The site is updated and the chatbot keeps quoting last quarter's page

Retraining is a console chore Someone remembers to re-scrape, or nobody does

No check that it worked The source shows as trained and nobody asks the bot the question

With FlowRunner

Publishing drives retraining A published page triggers the re-scrape of exactly the source that holds it

Indexing waited on properly The flow holds until the source reaches success rather than assuming it

Verified against real questions Canary questions are asked and their citations compared before anyone calls it done

Use Case Scenarios

A published page that retrains its own chatbot

The documentation site publishes and fires a webhook. The flow matches the published URL against List Data Sources, finds the link source that already holds that page, and calls Retrain URL Data Sources on its UUID rather than adding a duplicate source. It polls until the status reaches success rather than assuming the scrape finished, because the answer to a canary question asked mid embedding is meaningless. Then it verifies: Create Session, three questions the page is meant to answer, List Messages to read the citations and confirm the answers now cite the retrained source. Only after that does it post the result. The corpus and the site stay in step because publishing is what moves them, not somebody's recurring calendar reminder.

Transcripts out of the platform in one call

Support wants a record of the conversations the bot handled and a sample scored for quality. On New Session fires with the session UUID and the lead form details from the meta object. When the conversation goes quiet, the flow calls Get Session Transcript, which returns the whole exchange as plain alternating User and Assistant lines with no structure to unpick. That text goes straight into a Zendesk ticket, or into a summarisation step, or into a scoring prompt. Conversations the scorer flags are surfaced to a QA reviewer with the transcript in the message, and the reviewer decides whether the answer was actually wrong or merely blunt. Nobody parses a message array to do any of it.

Making a chatbot reachable, on purpose

Visibility on a GPT-trainer chatbot has three settings and they are not interchangeable. Private keeps the bot inside the account. Public allows link sharing on gpt-trainer.com as well as website embedding. Hybrid allows embedding only. Update Chatbot changes that with one field, which means an internal bot trained on internal runbooks can become publicly reachable through a single automated write. So a flow that manages visibility treats the move to Public as a publishing act rather than a settings change. It reads the current value with Get Chatbot, enumerates what the bot is trained on with List Data Sources, and asks the owner directly: this bot answers from 41 sources including the internal pricing sheet, and Public makes it link shareable outside the account. A person answers that question. The flow can move a bot to Private on its own, because tightening reach breaks nothing that matters.

Human-in-Loop Highlight

Delete Multiple Data Sources is the gate, and the reason is a property of the call itself rather than the size of the batch.

The operation takes a list of source UUIDs and deletes them, and those sources may belong to any chatbot on the account. There is no chatbot parameter scoping the request. That is convenient when you are clearing out a batch of failed uploads and inconvenient in exactly one situation: when a cleanup flow is handed a list of UUIDs assembled from somewhere else. A stale export, a spreadsheet, a previous run's output. The call will happily reach into a production chatbot the flow was never pointed at and strip training out of it, permanently, and the only symptom afterwards is that a bot which used to answer a question stops being able to.

So a cleanup flow resolves the blast radius before it asks for anything. It walks List Data Sources per chatbot to map every UUID back to its owning bot, groups the proposed deletions by chatbot, and posts what it found: "Cleanup proposes deleting 63 training sources. 58 belong to the archived onboarding bot as expected. 5 belong to the live customer support bot, including the shipping policy page and two files still being cited in answers this week. Delete the 58, delete all 63, or stop?" The owner picks. Delete Multiple Messages gets the same treatment for the same structural reason: the messages may come from different sessions and every deletion is irreversible, so a redaction job that thinks it is scrubbing one test session can quietly redact a real customer's transcript. The agent does the mapping, which is the part it is genuinely better at. The person owns the deletion, which is the part that cannot be taken back. That split is what automation exceptions are for.

Agent processes routinely

Detects exception requiring judgment

Clear match Continues automatically

Ambiguous Routes to human via preferred channel

Human decides

Agent resumes with decision

Agent Capabilities

30 actions

Chatbots

Create Chatbot Creates a chatbot and returns the UUID every other operation needs. It starts with no training data, and visibility controls reach: Private keeps it inside the account, Public allows link sharing on gpt-trainer.com plus embedding, and Hybrid allows embedding only.
List Chatbots Returns every chatbot on the account with its UUID, timestamps and meta settings such as visibility, rate limit and citation display. Used to discover the UUIDs the session, agent and data source operations need.
Get Chatbot Retrieves one chatbot by UUID including its meta object holding visibility, rate limit configuration and the citation setting. Read before applying a partial update.
Update Chatbot Updates a chatbot's meta settings, changing only the values supplied, so it can rename, switch visibility, adjust the per device rate limit or toggle citations independently. Rate limit messages and seconds must be supplied together.
Delete Chatbot Permanently deletes a chatbot together with its agents, training sources, sessions and message history. Any embed or shared link pointing at it stops working immediately.

Agents

Create Agent Adds an agent to a chatbot. User-Facing agents answer visitors from the source library, Background agents monitor conversations and apply tags, Human Escalation routes the visitor to a person, Pre-Canned returns a fixed reply and Spam Defense filters abuse. Only one Human Escalation and one Spam Defense agent are allowed per chatbot, and new agents are created disabled.
List Agents Returns every agent on a chatbot with its UUID, type, prompt, enabled flag, meta settings, the data source UUIDs it is restricted to, its escalation settings and any tool functions or variables it declares.
Update Agent Updates an agent and is the operation that enables it, since agents are created disabled. Data Source UUIDs restrict a User-Facing agent to specific sources, or Use All Sources gives it the whole library.
Delete Agent Permanently removes an agent from its chatbot. Training sources are untouched, but any behaviour that depended on the agent, such as escalation routing or conversation tagging, stops immediately.

Data Sources

Upload Data Source File Downloads a file from a URL, including a FlowRunner file URL, and uploads it as a training source. Chunking, embedding and indexing are asynchronous, so the status is polled until it reaches success. A reference source link controls the URL the chatbot links to when citing the document.
Create URL Data Source Adds a web page as a training source. GPT-trainer scrapes it and then chunks and embeds asynchronously, so the returned status is usually an in-progress value before it reaches success.
Create Q&A Data Source Adds a hand written question and answer pair, which is the most reliable way to make a chatbot answer a specific question exactly as intended. It is indexed alongside files and scraped pages and can be cited like any other source.
List Data Sources Returns every training source on a chatbot with its UUID, generated title, type, indexing status, token count, file size and metadata. Statuses separate in-progress values from success and from error values, including distinct storage and token limit errors.
Update Data Source Title Renames a training source. The title is generated automatically on ingest and is what the chatbot displays when it cites that source, so a clearer title produces a more readable citation for end users.
Delete Data Source Permanently removes one training source and its embeddings. Answers that relied on it stop being available immediately and the tokens it consumed are released back to the plan.
Delete Multiple Data Sources Permanently removes several training sources in one call. The sources may belong to any chatbot on the account and every deletion is irreversible, which is why flows resolve each UUID to its owning chatbot before proposing the batch.
Retrain URL Data Sources Re-scrapes link sources so the chatbot picks up current page content, which is the standard way to keep it in step with a documentation site or price list. Only link sources can be retrained, and the response returns a per UUID status rather than failing the whole call.

Source Tags

Create Source Tag Creates a coloured tag for organising training sources and optionally assigns sources to it immediately. The usual way to group a large library by product line, language or department.
List Source Tags Returns every tag on a chatbot with its UUID, name, colour and current source membership. Used to see how a library is organised or to find the tag UUID for an update.
Update Source Tag Updates a tag's name, colour or membership. The source UUID list replaces the tag's current contents rather than adding to it, so every source that should remain must be included.
Delete Source Tag Deletes a tag from its chatbot. Only the grouping is removed; the tagged sources remain and continue to be used for answers.

Sessions

Create Session Opens a chat session and returns its UUID. A session holds the conversation, keeps message order and preserves context between turns, so one must exist before Send Message. Initial Variables seeds it with data such as a customer name or account tier that agents and tools can reference.
List Sessions Returns the sessions of a chatbot, optionally narrowed to a creation date range, each with its UUID, timestamps and a meta object holding lead form details and verified auth data where identity verification with data saving is enabled.
Get Session Retrieves one session by UUID with its timestamps and meta object, which is the operation to call when a workflow needs to know who a conversation belongs to.
Delete Session Permanently deletes a session and the messages it contains. Used to honour data deletion requests or clear test conversations; the chatbot and its training sources are unaffected.

Messages and Transcripts

Send Message Sends a query to a session and returns the chatbot's answer. GPT-trainer streams the answer and this operation waits for the stream to finish, returning the accumulated reply as text. Reuse the session UUID for follow ups so context is kept.
List Messages Returns every message in a session as structured records holding the query, the response, the finish reason, the citation data produced when citations are enabled, and message metadata. The operation to call after Send Message when citations matter.
Get Session Transcript Retrieves the whole conversation as a plain text transcript with alternating User and Assistant lines. The quickest way to hand a full conversation to a summariser, a support ticket or an email without post processing.
Delete Message Permanently deletes one message from a session, removing both the visitor query and the chatbot response it recorded. The session and its remaining messages are untouched.
Delete Multiple Messages Permanently deletes several messages in one call, which is the efficient way to redact a run of exchanges or clear test traffic. The messages may come from different sessions and every deletion is irreversible.

Triggers

1 triggers

Event Triggers

On New Session Fires when a new chat session is opened on the watched chatbot, whichever channel it came from: the embedded widget, a shared link or this integration. The emitted session carries its UUID plus the meta object holding lead form details, which makes it the natural entry point for routing new conversations into a CRM or alerting a team. Follow it with List Messages or Get Session Transcript to read what was actually said. The first cycle records existing sessions without firing, so history is never replayed.

Frequently Asked Questions

What can FlowRunner do with GPT-trainer?

FlowRunner agents can run Create Chatbot, List Chatbots, and Get Chatbot in GPT-trainer, plus 27 more actions.

Does connecting GPT-trainer to FlowRunner require OAuth?

No. GPT-trainer connects to FlowRunner with an API key, no OAuth flow required.

Can GPT-trainer trigger a FlowRunner workflow automatically?

Yes. GPT-trainer supports 1 trigger that can start a FlowRunner workflow automatically.

Start building with GPT-trainer

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

Start Building Free Talk to us