I just had an interesting conversation with ChatGPT talking about assistants and the assistants API in particular. Recall that the Assistants API remains in beta. I also note that the Completions API is getting new features, such as memory.
In that conversation, we discussed RAG and the file_search tool in particular. ChatGPT told me that file_search is based on the FileSearch API which is now deprecated. It is recommended that search now be performed (in a Completions context) using Embeddings. But it appears that the Assistants API still uses file_search and it still uses the deprecated search.
This leads me to speculate that perhaps the roadmap (unstated publicly as far as I can tell) may be to eventually eliminate the Assistants API and simply keep adding features to the Completion API (like RAG) until it obviates the need for a separate Assistants API.
This is pure speculation on my part. But if this is true, it would be VERY helpful to hear something about this on an official basis from OpenAI. We’ve been waiting for Assistants API to come out of beta for months now. If it is never going to come out, I’d love to have a heads-up.
Thanks for sharing your findings.
I have been waiting for simple fixes in the Assistants API, like making it work with image inputs but getting no traction from OpenAi for like a quarter now which suggests no-one is taking it seriously anymore.
I loved the memory of Assistants, I hate using langchain, it’s just so much overdone.
If you read the Model Spec all they talk about is Assistants. I’m sure it’s the foundation for GPT-5 and they’re just getting it ready for that release this year.
I checked on this again. There was an old Search API that has been deprecated. It was part of a package of stuff including Classification stuff, etc.
But I can’t find anything (in spite of that GPT conversation I had) that says that file_search is implemented on that deprecated Search API. So I don’t want to be guilty of spreading rumors.
Perhaps @jim is right and Assistants is alive and well and part of gpt-5. Let’s hope so.
Assistants is a specific meaning used in chat completions and the internal ChatML containers for specifically describing the “assistant” separator that that acts as a prompt for language generation from a different entity, instead of having the internal prompt be “AI” or “You:”. Before DevDay.
Assistants, within which you create assistants and receive an assistants message where an assistants prompt after messages is fulfilled by an assistant’s instructions and then you get the assistant role message from messages threads which works like an API GPT then fulfilled by but completely different than gpt-4-turbo…was clearly meant to be nomenclusterfu. The code name was “gizmos”.
v1 had “retrieval” to empty your credits out of your account with how much internal calling and reading it could do. They were trying to write a “deep research” on your documents, you can see (but not see the costs until you got the daily billing).
Not just a document store search, but clicking into a document, then reading the text, then searching the document, then searching the document, then searching the document, then marking a text passage, then going back, then selecting another document…
The full extent of methods and instructions was hidden until within the iterations, and the instructions depend on the date and the stage. Preserved her for self-documenting posterity:
(instructions)
# Tools
## myfiles_browser
You have the tool `myfiles_browser` with these functions:
## myfiles_browser
You have the tool `myfiles_browser` with these functions:
`search(query: str)` Runs a query over the file(s) uploaded in the current conversation and displays the results.
`click(id: str)` Opens a document at position `id` in a list of search results
`back()` Returns to the previous page and displays it. Use it to navigate back to search results after clicking into a result.
`scroll(amt: int)` Scrolls up or down in the open page by the given amount.
`open_url(url: str)` Opens the document with the ID `url` and displays it. URL must be a file ID (typically a UUID), not a path.
`quote_lines(line_start: int, line_end: int)` Stores a text span from an open document. Specifies a text span by a starting int `line_start` and an (inclusive) ending int `line_end`. To quote a single line, use `line_start` = `line_end`.
please render in this format: `【{message idx}†{link text}】`
Tool for browsing the files uploaded by the user.
Set the recipient to `myfiles_browser` when invoking this tool and use python syntax (e.g. search('query')). "Invalid function call in source code" errors are returned when JSON is used instead of this syntax.
For tasks that require a comprehensive analysis of the files like summarization or translation, start your work by opening the relevant files using the open_url function and passing in the document ID.
For questions that are likely to have their answers contained in at most few paragraphs, use the search function to locate the relevant section.
Think carefully about how the information you find relates to the user's request. Respond as soon as you find information that clearly answers the request. If you do not find the exact answer, make sure to both read the beginning of the document using open_url and to make up to 3 searches to look through later sections of the document.
A bit later version, and how ChatGPT GPTs injected files with messaging regardless of relevance:
I agree with this post, the Assistants API is still in Beta. There is no roadmaps and all we can see that some of the basic issues of Assistants API does not get fixed and the Chat Completions API just keeps getting more features and updates. At this point I would not build anything to production with Assistants API.
When I started noodling with openAI a year back I got the impression the assistants API and all related (thread, messages, runs) would be the main “standard” moving forward so I invested quite a bit in that.
Just got my impressions reset a month back that perhaps chat completions was going to be the main thrust moving forward so I ported over to that api. Apart from 1 intermittant 500 response (and and older version of tools functions json schema) it’s faster and more featureful for me.
I wish I had realized this sooner. When a company has limited resources to build APIs it’s always better to be tracking what is considered the “primary” – so this would be chat completions (aka the “Vanilla” API) ?
In one of the releases about Responses it specifically indicated that they will sunset Assistants in 2026. Presumably they are also talking about Assistants v2 (beta). I am curious to know if anyone here has found if Responses is capable of the persistent memory (threading) that Assistants has. I love that the web search feature models exist for Responses, but not for Assistants V2, unfortunately.
as far as I understand (and as stated in the docs), you create “responses” based on a conversational history that you either provide or that refers to a previous “response” object
So unlike “threads”, you don’t reuse the same “response” over and over, you create new ones and their history “grows”… so the last “response” object you created contains the entire conversational history nonetheless …
“Response” objects however are saved only for 30 days though (or not saved at all if you want)… I hope they’ll make this more persistant and let the user decide if the response object must expire.
The new Responses API makes it a lot easier to branch a conversation. In comparison, it takes a lot more effort to do the same with Thread manipulation in Assistants API.
I’ve only just seen this and finished a change from Assistant API to chat completions. A main driver was speed – any idea how responsive (timewise) this API is compared with chat completions. I found assistant API 3x slower than chat completions.