AI Assistants Compared: Custom Knowledge Bases and Private Data Integration
Vendor pages document different upload limits, data-use settings, and cost structures, so compare each assistant with the same private files and questions before choosing.
Sources checked 2 Oct 2026
ChatGPT, Claude, and Gemini document file-based work through different units: ChatGPT Projects and custom GPT knowledge, Claude Projects, and Gemini uploads or Drive-connected files. Their pages settle documented limits and settings, not which assistant will retrieve private material best; Chat Picker has not tested them, so you should compare them with the same controlled corpus.
Know these limits before you start
Start with ingestion, not model choice. OpenAI's File Uploads FAQ gives a 512MB hard limit per conversation file, a 2M-token cap for text and document files, and 3 uploads per day for Free users, as read on October 1, 2026. Anthropic's upload page lists 500MB per chat file but 30MB per project file; it analyzes PDF visual elements through 100 pages and text only from 101 to 1,000 pages, as read on October 1, 2026. Google's file-upload page says a prompt can include up to 10 supported files, subject to availability, with videos up to 2GB and other files up to 100MB, as read on October 1, 2026. These are app rules, not proof of retrieval quality or API ingestion.
Private-data settings differ. OpenAI's pricing and Projects pages say an opt-out is available on Free, Go, Plus, and Pro, while business workspace content is not used for training by default, as read on October 1 and 2, 2026. Anthropic's pricing page says Free, Pro, and Max have an opt-out and Team is not trained on by default, as read on October 1, 2026. The Google plans page read on October 1, 2026 does not state a chat-training setting; leave it unresolved.
Each vendor warns against relying on output: OpenAI says ChatGPT can be incorrect or misleading, Claude should not be the only source of truth, and Gemini is probabilistic and may sometimes violate its harmful-inaccuracy guidelines, as read on October 1, 2026. OpenAI prohibits unauthorized aggregation of private or sensitive information, while Anthropic requires qualified professional review for covered high-risk advice, as read on the same date.
The test plan
Use synthetic files until your policy and access review is complete. Keep one log row per file or question.
-
Build a controlled corpus. Input: make TXT, PDF, XLSX, and PNG copies containing: “For the test account, renewal requires written approval, and the billing team handles billing questions.” Add an obsolete note saying the support team handles billing. Instruction:
List every file you can read and its title; do not answer yet.Observe omissions or parse errors. Log the filename, format, size, and source location. -
Check ingestion. Input: use the same files in a fresh chat or project. Instruction:
Use only the uploaded files. State whether each file was accepted and readable; do not infer missing text.Observe rejections, warnings, truncation, and re-upload requests. Log accepted or rejected status, warnings, interface, plan, and run date. -
Test retrieval. Input: use the same corpus, including a new chat within the same project where persistence is offered. Instruction:
Answer only from the corpus. State the test account's approval rule and billing team. For each answer, give the filename and exact supporting sentence. Then say whether the support note conflicts. If evidence is absent, write “Not found.”Observe the source, quote, conflict detection, unsupported additions, and whether project knowledge persisted. Log each item separately. -
Test access boundaries. Input: use the files and an authorized connected source under owner, editor, and viewer roles if available. Instruction to the viewer:
List the project instructions and knowledge files available to your account, then state whether you can edit each item. Do not access anyone else's material.Observe visibility and blocked edits; never try to bypass controls. Log role, connector or admin restrictions, and unresolved role-based access, encryption, or data-isolation terms. -
Check API and portability. Keep originals and a manifest outside the platform. If using an API, submit the corpus with:
Index only the supplied test corpus. Return each filename, matching quote, and stable citation reference. Do not add absent facts.Observe accepted formats, conversions, file IDs, citations, embeddings, fine-tuning access, export or download, and steps needed to rebuild the index elsewhere. Log each capability as available, unavailable, or unverified. -
Measure overhead. Run the same retrieval set on the selected consumer plan and API separately. Log plan fees, tokens, tool charges, storage, engineering time, and any total shown by the vendor. OpenAI's API pricing lists file-search storage at $0.10 per GB per day with 1GB free, while Anthropic's pricing documentation lists web search at $10 per 1,000 searches plus standard content-token charges, as read on October 1, 2026. Chat Picker could not verify a knowledge-base storage figure from Google's developer pricing page, as read on the same date. Calculate cost per completed, source-backed answer; do not mix consumer and API figures.
How to read your results
Set hard gates first: data setting, permitted content, access, retention, export, and vendor policy. Remove any option that fails one. For the rest, compare your logs by failure type: rejected file, missing source, missed conflict, unsupported statement, unexpected access, portability work, and cost.
Re-run failures in a fresh chat to separate ingestion from retrieval. Record the account, plan, interface, and conditions instead of generalizing to every model. Choose only among options that pass your gates; Chat Picker provides no test score or ranking.
Where the plans differ
Use the three-way comparison matrix rows for context window, file uploads, projects and custom knowledge, training controls, access and permissions, retention and export, embeddings and fine-tuning