Upload Document to Chat (Ad-Hoc context)

Overview

Flow ID: upload-document-to-chat
Category: Chat Interactions
Estimated Duration: 10-30 seconds
User Role: All Users
Complexity: Simple

Purpose: Quickly add a single file (PDF, TXT, etc.) to the current chat context without creating a permanent Dataset. This is useful for “talk to this PDF” scenarios where the document is transient or one-off.

Trigger

What initiates this flow:

• User manually initiates

Specific trigger: Clicking the Paperclip Icon in the chat input bar.

Prerequisites

Before starting, users must have:

• Supported file types (PDF, DOCX, TXT, CSV)
• File size under limit (usually ~10MB or token equivalent)

User Intent Analysis

Primary Intent

“I just want to ask a question about this specific file right now, I don’t want to manage a database.”

Step-by-Step Flow

Main Path (Happy Path)

Step 1: Upload File

• User Action: Click Paperclip Icon > Select File.
• System Response:
  • File appears as “chip” in input area.
  • Progress bar (uploading/parsing).

Step 2: Processing (Fast)

• System Action:
  • Extract text.
  • (Optional) Rough chunking or full-text injection depending on size.
  • Note: This usually bypasses the heavy vector embedding job pipeline for immediate availability in RAM.

Step 3: Ask Question

• User Action: Type “Summarize this” or “What are the main points?”
• System Response: AI answers using the uploaded file’s content.

Step 4: Persistence Check

• Condition: Does this file stick around?
• Behavior: Usually Session-Scoped. If the user leaves the chat, the file context is lost unless explicitly saved to a dataset.

Error States & Recovery

Error 1: File Too Large (Token Limit)

Cause: Uploading a 500-page book into a context window of 8k tokens
User Experience: “File too large for context window.”
Recovery: System suggests adding it to a Dataset (Blockify Job) instead for full indexing.

Error 2: Unreadable Text

Cause: Scanned PDF (Image) without OCR
User Experience: AI hallucinates or says “It looks empty.”
Recovery: User must OCR file externally first.

Pain Points & Friction

1. Token Accounting: Users don’t understand that a 5MB text file might consume their entire context budget, confusing the AI.
   • Mitigation: “Token Usage” bar clearly showing file impact.

Design Considerations

• Visual Feedback: Distinct icon for “File in Context”.
• Interaction: Allow clicking the file chip to “Remove” it from pending message.

Related Flows

• Rag Enabled Chat - The robust, persistent version
• Upload Files to Job - Batch processing

Technical References

• src/components/chat/file-uploader.js
• src/utils/text-extractor.js