ChatGPT Projects allow users to upload files, but they do not allow the model to deterministically consult those files during reasoning. As a result, ChatGPT cannot reliably function as a long-running knowledge partner or true reference for projects that depend on canonical user-supplied data.
This limitation is not about hallucinations, safety failures, or model reasoning quality. It is a missing product capability: indexed, queryable access to user-provided project files.
Many advanced use cases depend on a stable corpus of authoritative material supplied by the user – hobbyist technical references, association documents, club rules, saved plans or recipes, and hundreds of other cases that people encounter in their everyday life. In these cases, users do not want probabilistic recall or conversational memory. They want answers grounded explicitly in the data they have provided for that particular project, or a clear statement that the data needed to answer a question is not present.
Today, even when users upload complete project files and explicitly correct errors, ChatGPT cannot reliably check those files when answering future questions. This leads to silent contradictions with uploaded facts, difficulty correcting the model when it is wrong, and an erosion of trust in long-running projects. The only reliable workaround is for users to manually extract, paste, and re-introduce context into prompts, which defeats the purpose of a conversational assistant.
This request is not about web browsing, live external data, or mutable third-party sources. It is not about asserting truth over the outside world. It is specifically about user-owned data that the user has chosen to upload and treat as authoritative within a project. Nearly everyone has projects that have their own related corpus of information, and they should be able to create a project and incorporate their own data into the project as more than simply additional prompt data.
A minimal, safety-aligned solution seems achievable without introducing new risk. Specific project files could be identified by user as project reference files - these are to be indexed once at upload, using deterministic text extraction and chunking, and treated as immutable thereafter. During reasoning, the model could query this index within the project scope, with clear rules: if relevant data is found, prefer it over conversational memory. Optionally, the UI could indicate when an answer is grounded in indexed project files or surface which files were consulted.
This approach does not require live access, external calls, or mutable state. It aligns naturally with existing retrieval-augmented generation patterns, but keeps the entire system confined to user-supplied, read-only data.
The absence of this capability is becoming more visible as users grow more sophisticated. Many now expect AI tools to combine strong reasoning with reliable access to their own data, rather than forcing an either-or choice. Competing tools already support deterministic interaction with user datasets through mechanisms such as MCP or direct codebase indexing, and users are beginning to migrate not because ChatGPT reasons poorly, but because it cannot reliably ground its reasoning in supplied information.
In practice, users can already solve this problem externally by exporting their project data, indexing & searching it locally, and manually feeding excerpts back into ChatGPT. That workaround demonstrates that the capability is feasible – but also highlights that it belongs inside the product rather than in user-built scaffolding.
ChatGPT’s reasoning capabilities are already strong. What is missing is a reliable substrate for user-provided truth. Indexed, immutable project files would unlock a wide range of serious, long-running uses and would meaningfully improve trust, correctness, and user experience without introducing new safety concerns.
I believe this is a foundational capability rather than a niche feature, and I hope it can be considered as part of the future evolution of Projects.