Knowledge attachment

A knowledge attachment is the file uploaded to a custom persona (ChatGPT GPT, Claude Project, Gemini Gem) that the LLM retrieves over at inference — turning custom personas into mini RAG apps without writing retrieval code.

Knowledge attachments make custom-persona platforms RAG-capable out of the box: upload PDFs, docs, CSVs, the platform indexes them, retrieval happens automatically per turn. Limits vary by platform: ChatGPT GPTs accept up to 20 files (~2MB each), Claude Projects accepts ~100MB, Gemini Gems take files via the Gemini app. Production gotchas: retrieval quality varies (semantic vs keyword), attachment-content-extraction can mangle complex PDFs (tables, scanned docs), no fine-grained access control (per-attachment auth), updates require re-uploading. For high-quality RAG production teams still build their own pipeline. For internal team knowledge bases knowledge attachments are 'good enough' to ship in a day.

When to use knowledge attachment

• Quick internal-knowledge bots.
• Small-scale RAG personas without backend code.

Common mistakes

• Trusting attachment retrieval quality on complex PDFs — tables and scanned docs degrade.
• Treating uploaded files as access-controlled — anyone with the persona link can extract them.

FAQ

Related terms

• Retrieval-augmented generation (RAG) — Retrieval-augmented generation (RAG) injects relevant documents into the prompt at query time so the model answers from your data instead of its training memory.
• Vector database — A vector database stores embeddings and performs approximate nearest-neighbor search at scale, the persistence layer behind RAG and semantic search.
• Custom instruction — Custom instructions are user-level or persona-level system prompts that personalize an LLM's behavior across sessions — ChatGPT Custom Instructions, Claude Style / Projects, Gemini Gems system prompts, Poe Bot definitions are 2026 implementations.

Last updated: 2026-06-01. Raw markdown: https://promtable.com/glossary/knowledge-attachment.md.