OpenAI

OpenAI models. Generate text and images, compute embeddings, moderate content, and manage files, fine-tuning, and vector stores.

Measured routing accuracy

54 golden cases replayed through the router over the whole index: measured over corpus ea4f12ad2948 (65 toolkits, 2283 tools indexed and 13 declared uncallable), 27 cases written by hand and 27 cases from the paraphrase pass. A case counts as top-1 when its gold tool ranked first and top-8 when it reached the slate at all.

5 of the 32 action tools listed above are declared uncallable, so they are indexed nowhere and answer none of these cases. The counts here are scored over the other 27.

The sweep is offline: the reranker is a deterministic identity fake that returns candidates in the order retrieval produced them, so top-1 measures retrieval order rather than a reranked slate. just eval-live measures the same cases through the live reranker.

Authentication

Connect an entity with ConnectionsService.InitiateConnection, naming this toolkit's slug. Credentials stay in the connections vault; callers hold connected-account ids only.

api_key

Submit the key with ConnectionsService rather than putting it in a request; it is sealed at rest and never returned.

Tools

32 action tools. The catalog-wide slug is openai.<tool>, which is what search_tools returns and call_tool takes.

openai.create_chat_completion

Generate a chat completion from an OpenAI GPT model (gpt-4o, gpt-4, gpt-3.5-turbo, or another chat model by name). Use when the user wants an AI-written reply, summary, draft, or answer produced from a prompt or conversation, rather than a saved template. Takes a system prompt plus message history; no function/tool calling in this schema.

Class write (writes, no confirmation needed). No scopes beyond the connection's defaults. Calls POST /chat/completions.

Arguments:

Result fields (the payload is trimmed to these before it reaches the model):

Also retrieved by: "ask gpt to write a reply for this", "have the model draft an answer to this question", "get an ai-written summary of this text", "run this prompt through gpt-4o and see what comes back", "generate a response from the language model for this conversation".

openai.create_image

Generate a brand-new image from a text prompt with an OpenAI image model (gpt-image-1 or dall-e-3). Use when the user wants original artwork or a picture created purely from a description. To modify an existing image instead of creating a new one, this toolkit cannot help (edit_image and create_image_variation are not callable here).

Class write (writes, no confirmation needed). No scopes beyond the connection's defaults. Calls POST /images/generations.

Arguments:

Result fields (the payload is trimmed to these before it reaches the model):

Also retrieved by: "make a picture of a robot riding a bicycle", "draw something from this description", "generate artwork of a sunset over mountains", "turn this idea into a piece of ai artwork", "render an original picture of a cat wearing a hat".

openai.create_embedding

Compute an OpenAI text embedding vector for one or more strings, turning text into numbers for search or similarity comparison. Use when the user wants to embed, vectorize, or compare text semantically, not when they want the model to write a reply.

Class write (writes, no confirmation needed). No scopes beyond the connection's defaults. Calls POST /embeddings.

Arguments:

Result fields (the payload is trimmed to these before it reaches the model):

Also retrieved by: "turn this sentence into a vector for search", "compute a numeric representation of this text", "get the vector for this string so i can compare it to others", "vectorize this paragraph for semantic search".

openai.create_moderation

Check whether text is flagged as unsafe by OpenAI's moderation model, across categories like hate, harassment, self-harm, sexual content, and violence. Use before showing user-submitted or model-generated text to others, or before sending it on to another tool. Read-only: it classifies text and creates nothing on the account.

Class read (reads only). No scopes beyond the connection's defaults. Calls POST /moderations.

Arguments:

Result fields (the payload is trimmed to these before it reaches the model):

Also retrieved by: "check whether this text is safe to show people", "is this message flagged for anything", "scan this for hate speech or harassment before I post it", "run a safety check on this user submission", "flag anything unsafe in what the model just wrote".

openai.create_transcription

Transcribe an audio file to text with an OpenAI speech-to-text model (Whisper). Use when the user wants a written transcript of an audio recording in its original language.

Class write (writes, no confirmation needed). No scopes beyond the connection's defaults. Calls POST /audio/transcriptions.

Arguments:

Result fields (the payload is trimmed to these before it reaches the model):

Also retrieved by: "turn this recording into written words", "what did they say in this voice memo", "give me a text version of this podcast clip".

openai.create_translation

Translate an audio file directly into English text with an OpenAI speech-to-text model (Whisper). Use when the user has audio in another language and wants an English transcript, not a transcript in the original language.

Class write (writes, no confirmation needed). No scopes beyond the connection's defaults. Calls POST /audio/translations.

Arguments:

Result fields (the payload is trimmed to these before it reaches the model):

Also retrieved by: "turn this foreign-language recording into an english transcript", "what is being said in this clip, in english", "give me an english write-up of this spanish voicemail".

openai.edit_image

Edit or inpaint part of an existing image from a text prompt and an optional mask, with an OpenAI image model. Use when the user wants to change one area of a picture they already have rather than generate a new one from scratch.

Class write (writes, no confirmation needed). No scopes beyond the connection's defaults. Calls POST /images/edits.

Arguments:

Result fields (the payload is trimmed to these before it reaches the model):

Also retrieved by: "paint over just this part of the picture", "swap out the background in this photo using a mask", "change one area of this picture, keep the rest the same".

openai.create_image_variation

Generate variations on an existing image without a text prompt, with an OpenAI image model. Use when the user wants a few alternate takes on a picture they already have, rather than a specific edit or a brand-new image.

Class write (writes, no confirmation needed). No scopes beyond the connection's defaults. Calls POST /images/variations.

Arguments:

Result fields (the payload is trimmed to these before it reaches the model):

Also retrieved by: "give me a few different takes on this picture", "remix this image into some alternates", "show me some fresh spins on this existing artwork".

openai.list_models

List every OpenAI model available to the account, base models and the account's own fine-tuned models alike. Use when the user asks what models they can use or wants a fine-tuned model's exact id.

Class read (reads only). No scopes beyond the connection's defaults. Calls GET /models.

Takes no arguments.

Result fields (the payload is trimmed to these before it reaches the model):

Also retrieved by: "what language models can I use here", "show me every model available on this account", "what's the exact id of my custom trained model".

openai.retrieve_model

Get one OpenAI model's details by id, base or fine-tuned. Use when the user already has a model id and wants to confirm it exists or see who owns it, rather than browsing the full list.

Class read (reads only). No scopes beyond the connection's defaults. Calls GET /models/{{params.model}}.

Arguments:

Result fields (the payload is trimmed to these before it reaches the model):

Also retrieved by: "look up the details for this specific model id", "confirm this model still exists and who owns it", "check the info on one particular model".

openai.delete_fine_tuned_model

Permanently delete one of the account's own OpenAI fine-tuned models by id. Use when the user wants to remove a fine-tuned model they trained. Only fine-tuned models can be deleted this way; base models cannot. Irreversible.

Class destructive (held by the mutation gate until the call's own arguments carry "confirm": true). No scopes beyond the connection's defaults. Calls DELETE /models/{{params.model}}.

Arguments:

Result fields (the payload is trimmed to these before it reaches the model):

Also retrieved by: "get rid of this custom trained model for good", "remove the fine-tuned model I no longer need", "wipe out that custom model from the account".

openai.upload_file

Upload a file to OpenAI's storage for use as fine-tuning training data, a batch job's input, or a vector store document. Use when the user wants to add a new file to the account rather than reference one that's already there.

Class write (writes, no confirmation needed). No scopes beyond the connection's defaults. Calls POST /files.

Arguments:

Result fields (the payload is trimmed to these before it reaches the model):

Also retrieved by: "add this document to the account's storage", "get this dataset uploaded so I can use it later", "push this file up for training or search".

openai.list_files

List files uploaded to the OpenAI account, optionally filtered by purpose (fine-tune, batch, vision). Use when the user asks what files they've uploaded or wants a file's id before referencing it in another call.

Class read (reads only). No scopes beyond the connection's defaults. Calls GET /files.

Arguments:

Result fields (the payload is trimmed to these before it reaches the model):

Also retrieved by: "what files have I uploaded so far", "show me everything sitting in storage right now", "find the id for that dataset I uploaded earlier".

openai.retrieve_file

Get one uploaded OpenAI file's metadata by id: name, size, purpose, and upload time. Use when the user already has a file id and wants its details, not the file's actual content or bytes.

Class read (reads only). No scopes beyond the connection's defaults. Calls GET /files/{{params.file_id}}.

Arguments:

Result fields (the payload is trimmed to these before it reaches the model):

Also retrieved by: "pull up the details on this uploaded document", "check the size and purpose of this stored file", "look up metadata for one item in storage".

openai.delete_file

Permanently delete an uploaded OpenAI file by id. Use when the user wants to remove a file from the account's storage. A file already in use by a fine-tuning job, batch, or vector store may still be referenced there after deletion fails or succeeds depending on OpenAI's own state; check first if that matters. Irreversible.

Class destructive (held by the mutation gate until the call's own arguments carry "confirm": true). No scopes beyond the connection's defaults. Calls DELETE /files/{{params.file_id}}.

Arguments:

Result fields (the payload is trimmed to these before it reaches the model):

Also retrieved by: "remove this document from storage for good", "clear out that old dataset I don't need anymore", "get rid of this uploaded file".

openai.create_fine_tuning_job

Start an OpenAI fine-tuning job that trains a custom model from an already-uploaded training file. Use when the user wants to fine-tune or customize a base model on their own data. The training file must already exist in OpenAI's storage (see list_files); this toolkit cannot upload one directly.

Class write (writes, no confirmation needed). No scopes beyond the connection's defaults. Calls POST /fine_tuning/jobs.

Arguments:

Result fields (the payload is trimmed to these before it reaches the model):

Also retrieved by: "kick off a custom training run on my dataset", "start fine-tuning a model with the data I uploaded", "train a custom version of a base model".

openai.list_fine_tuning_jobs

List the OpenAI fine-tuning jobs run on the account, newest first. Use when the user asks what fine-tunes they've kicked off or wants a job's id before checking on it or cancelling it.

Class read (reads only). No scopes beyond the connection's defaults. Calls GET /fine_tuning/jobs.

Arguments:

Result fields (the payload is trimmed to these before it reaches the model):

Also retrieved by: "what training runs have I started", "show me every custom-training job on this account", "find the id for that fine-tune I kicked off earlier".

openai.retrieve_fine_tuning_job

Get one OpenAI fine-tuning job's current status by id: whether it's queued, running, succeeded, or failed, and the resulting model name once it finishes. Use when the user already has a job id and wants to check on a specific fine-tune.

Class read (reads only). No scopes beyond the connection's defaults. Calls GET /fine_tuning/jobs/{{params.fine_tuning_job_id}}.

Arguments:

Result fields (the payload is trimmed to these before it reaches the model):

Also retrieved by: "how's my custom training run coming along", "check on the progress of one specific fine-tune", "has my model finished training yet".

openai.cancel_fine_tuning_job

Stop an OpenAI fine-tuning job that hasn't finished yet. Use when the user wants to cancel, kill, or abort a fine-tune that's still queued or running. Irreversible; a job that has already completed or failed cannot be cancelled.

Class destructive (held by the mutation gate until the call's own arguments carry "confirm": true). No scopes beyond the connection's defaults. Calls POST /fine_tuning/jobs/{{params.fine_tuning_job_id}}/cancel.

Arguments:

Result fields (the payload is trimmed to these before it reaches the model):

Also retrieved by: "stop this training run before it finishes", "kill that custom-training job, it's not needed anymore", "abort the fine-tune that's still going".

openai.list_fine_tuning_events

List the progress events (metrics, status changes) an OpenAI fine-tuning job has logged, oldest or newest first. Use when the user wants to see how a fine-tune is progressing in detail, beyond just its current status.

Class read (reads only). No scopes beyond the connection's defaults. Calls GET /fine_tuning/jobs/{{params.fine_tuning_job_id}}/events.

Arguments:

Result fields (the payload is trimmed to these before it reaches the model):

Also retrieved by: "show me the play-by-play of this training run", "what's been logged during this custom-training job", "give me the detailed progress log, not just the status".

openai.create_batch

Submit an OpenAI batch job: run many chat completion or embedding requests from one already-uploaded JSONL input file, at lower cost and with results delivered within a completion window rather than immediately. Use when the user wants to process a large set of prompts asynchronously instead of one at a time.

Class write (writes, no confirmation needed). No scopes beyond the connection's defaults. Calls POST /batches.

Arguments:

Result fields (the payload is trimmed to these before it reaches the model):

Also retrieved by: "run all these prompts overnight instead of one by one", "submit a big pile of requests as a bulk job", "process this large set of inputs asynchronously".

openai.retrieve_batch

Get one OpenAI batch job's current status by id: how many of its requests have completed or failed, and the output file id once it's done. Use when the user already has a batch id and wants to check on a specific bulk job.

Class read (reads only). No scopes beyond the connection's defaults. Calls GET /batches/{{params.batch_id}}.

Arguments:

Result fields (the payload is trimmed to these before it reaches the model):

Also retrieved by: "how far along is that bulk job I submitted", "check the progress on this large async run", "has the batch I kicked off finished yet".

openai.cancel_batch

Stop an OpenAI batch job that hasn't finished yet; in-flight requests are allowed to finish but no new ones start. Use when the user wants to cancel or abort a bulk job that's still in progress. Irreversible; a batch that has already completed cannot be cancelled.

Class destructive (held by the mutation gate until the call's own arguments carry "confirm": true). No scopes beyond the connection's defaults. Calls POST /batches/{{params.batch_id}}/cancel.

Arguments:

Result fields (the payload is trimmed to these before it reaches the model):

Also retrieved by: "stop that bulk job before it finishes", "abort the large async run that's still in progress", "kill the batch, we don't need those results anymore".

openai.list_batches

List the OpenAI batch jobs run on the account, newest first. Use when the user asks what bulk jobs they've submitted or wants a batch's id before checking on it or cancelling it.

Class read (reads only). No scopes beyond the connection's defaults. Calls GET /batches.

Arguments:

Result fields (the payload is trimmed to these before it reaches the model):

Also retrieved by: "what bulk jobs have I submitted", "show me every large async run on this account", "find the id for that batch I kicked off earlier".

openai.create_vector_store

Create a new OpenAI vector store: a searchable collection of files for retrieval. Use when the user wants to set up a new knowledge base or document collection for semantic search, not add a file to one that already exists.

Class write (writes, no confirmation needed). No scopes beyond the connection's defaults. Calls POST /vector_stores.

Arguments:

Result fields (the payload is trimmed to these before it reaches the model):

Also retrieved by: "set up a new knowledge base for search", "build a searchable collection out of some documents", "start a fresh document index for retrieval".

openai.list_vector_stores

List the OpenAI vector stores set up on the account. Use when the user asks what knowledge bases or document collections exist, or wants a vector store's id before adding files to it.

Class read (reads only). No scopes beyond the connection's defaults. Calls GET /vector_stores.

Arguments:

Result fields (the payload is trimmed to these before it reaches the model):

Also retrieved by: "what knowledge bases do I have set up", "show me every document collection on this account", "find the id for the search index I built earlier".

openai.retrieve_vector_store

Get one OpenAI vector store's details by id: its name, status, and how many files it holds. Use when the user already has a vector store id and wants its current state, not the full account list.

Class read (reads only). No scopes beyond the connection's defaults. Calls GET /vector_stores/{{params.vector_store_id}}.

Arguments:

Result fields (the payload is trimmed to these before it reaches the model):

Also retrieved by: "check how many files are in this knowledge base", "look up the state of one specific document collection", "is this search index still processing".

openai.update_vector_store

Rename an OpenAI vector store, or change when it expires. Use when the user wants to relabel a knowledge base or adjust its expiry, not change which files it contains.

Class write (writes, no confirmation needed). No scopes beyond the connection's defaults. Calls POST /vector_stores/{{params.vector_store_id}}.

Arguments:

Result fields (the payload is trimmed to these before it reaches the model):

Also retrieved by: "rename this knowledge base to something clearer", "change the expiry on this document collection", "relabel this search index".

openai.delete_vector_store

Permanently delete an OpenAI vector store by id. Use when the user wants to remove a knowledge base or document collection entirely. Removes the store's index, not the underlying uploaded files themselves. Irreversible.

Class destructive (held by the mutation gate until the call's own arguments carry "confirm": true). No scopes beyond the connection's defaults. Calls DELETE /vector_stores/{{params.vector_store_id}}.

Arguments:

Result fields (the payload is trimmed to these before it reaches the model):

Also retrieved by: "tear down this knowledge base entirely", "remove this document collection for good", "get rid of the whole search index".

openai.add_file_to_vector_store

Add an already-uploaded file to an OpenAI vector store, indexing it for retrieval. Use when the user wants to add a document to an existing knowledge base. The file must already exist in OpenAI's storage (see list_files); this toolkit cannot upload one directly.

Class write (writes, no confirmation needed). No scopes beyond the connection's defaults. Calls POST /vector_stores/{{params.vector_store_id}}/files.

Arguments:

Result fields (the payload is trimmed to these before it reaches the model):

Also retrieved by: "drop this document into the knowledge base", "index this file so it shows up in search", "add one more document to the collection".

openai.list_vector_store_files

List the files indexed in an OpenAI vector store, with each one's processing status. Use when the user asks what documents are in a knowledge base, or wants a file's status before searching against it.

Class read (reads only). No scopes beyond the connection's defaults. Calls GET /vector_stores/{{params.vector_store_id}}/files.

Arguments:

Result fields (the payload is trimmed to these before it reaches the model):

Also retrieved by: "what documents are indexed in this knowledge base", "show me every file in this search collection", "check whether this document has finished processing".

openai.remove_file_from_vector_store

Remove a file from an OpenAI vector store's index by file id. Use when the user wants a document taken out of a knowledge base's search results. Deletes the file's presence in this store's index only; the file itself stays in OpenAI's storage. Irreversible for this store's index.

Class destructive (held by the mutation gate until the call's own arguments carry "confirm": true). No scopes beyond the connection's defaults. Calls DELETE /vector_stores/{{params.vector_store_id}}/files/{{params.file_id}}.

Arguments:

Result fields (the payload is trimmed to these before it reaches the model):

Also retrieved by: "take this document out of the knowledge base", "stop this file from showing up in search results", "drop one document from the collection's index".