Intelligent Document Processing Skills

Intelligent Document Processing Skills

Overview

Use the File I/O skill set to let an iDialogue Agent process Salesforce Files for intelligent document processing (IDP). These skills give the agent a simple document workflow:

1. process_file prepares a file for downstream use.
2. read_file retrieves processed file content in bounded chunks.
3. write_file saves summaries, notes, or other agent-authored outputs back to the file workspace.

The processed representation is stored internally as markdown. Users do not need to ask for markdown explicitly.

Who This Is For

Salesforce administrators, solution architects, and developers who configure agents to review contracts, invoices, case attachments, intake forms, and other unstructured business documents.

Requirements

• Agent model must be GPT-5.4.
• Agent must have one or more File I/O skills enabled: process_file, read_file, write_file.
• The request context must provide the target Salesforce File version as a contentVersionId.
• Background execution is recommended for longer document-processing runs or multi-step document workflows.

What Each Skill Does

process_file

Use process_file when the agent needs to extract, normalize, or prepare a file for downstream use.

• Best used when the agent is seeing a file for the first time.
• Creates the canonical processed artifact for the file workspace.
• Returns status and preview metadata instead of the entire file body.

read_file

Use read_file when the agent needs to inspect, quote, summarize, or continue working with a processed file.

• Reads the canonical processed artifact or another saved markdown artifact.
• Returns bounded content so large documents do not flood the conversation.
• Can optionally ensure the file has already been processed before reading.

write_file

Use write_file when the agent needs to save a derived document, notes, or summary.

• Persists agent-authored markdown back to the file workspace.
• Useful for outputs such as summary.md, notes.md, or a refreshed file.md.
• Supports overwrite control so agents do not replace artifacts unintentionally.

Common Inputs

Admin Setup Checklist

1. Add the File I/O skills your agent needs: process_file, read_file, and optionally write_file.
2. Set the agent model to GPT-5.4.
3. Make sure your invocation flow can provide the target contentVersionId.
4. Update the system prompt so the agent knows when to process a file, when to read it, and when to save derived outputs.
5. Use background execution for long-running document jobs or multi-file workflows.

Prompt Guidance Snippet

Use this in your system or skill prompt to steer document-processing behavior:

When a user asks you to review, extract, analyze, or summarize a Salesforce File,
call process_file with the file's contentVersionId if the file has not been prepared yet.
Use read_file to inspect the processed content in bounded chunks.
Use write_file only when you need to save a summary, notes, or another derived document
back to the file workspace. Do not ask the user to provide markdown format.

Examples

Contract Review

1. A user asks the agent to review a contract attached to a Salesforce record.
2. The agent calls process_file with the contract contentVersionId.
3. The agent calls read_file to inspect the processed content.
4. The agent answers the user and optionally saves a summary.md artifact with write_file.

Invoice Intake

1. A user asks the agent to extract billing details from a PDF invoice.
2. The agent calls process_file to prepare the invoice.
3. The agent calls read_file to pull the relevant sections.
4. The agent returns the extracted values and can save reconciliation notes with write_file.

Policy or Case Attachment Follow-Up

1. A user asks for a concise summary of a policy document or case attachment.
2. The agent uses read_file if the file has already been processed.
3. If the file is not ready yet, the agent processes it first, then reads the relevant section.
4. The agent can save a reusable notes artifact for future follow-up.

Recommended Prompt Pattern

• Use process_file first when the document has not been normalized yet.
• Use read_file for inspection, quoting, summarization, or follow-up analysis.
• Use write_file only for derived outputs you want the agent to save.
• Treat the internal markdown representation as an implementation detail, not as end-user language.

Troubleshooting

• Agent is answering without using the file:
  • Confirm the agent has the File I/O skills enabled and the prompt tells it to call them.
• read_file cannot find content:
  • Process the file first, or configure the agent to use ensureProcessed=true when appropriate.
• write_file fails because the artifact already exists:
  • Choose a different artifact name or explicitly allow overwrite when replacement is intended.
• Long documents take multiple steps:
  • Use background execution and let the agent save summaries or notes instead of trying to return the entire document body in one response.