ZeroExport Ask Agent
Ask questions about the Word and PDF documents attached to any Salesforce record, answered from the document itself and cited by section.
ZeroExport Ask Agent is a separate installable app under the ZeroExport umbrella. It lets an Agentforce agent answer questions about the Word and PDF documents attached to a Salesforce record, reading the file at the moment the question is asked.
It installs and runs on its own. ZeroExport is not a prerequisite.
Current version: v0.3, which adds PDF reading alongside Word.
What it does
Open a record, ask a question in plain language:
"What's the target resolution for a Severity 1 issue?"
Severity 1 (Critical Outage) has a target resolution window of 4 hours, with a 15-minute first response, escalating to the VP of Engineering. (Section: 2. Targets - Enterprise_SLA_2026.docx)
Three things are happening there, and each one is the point.
- It read the document just now. Not an index built last week. If the file was uploaded a minute ago, the answer reflects it.
- It cited where the answer came from. Every fact is traceable to a section of a named document, so it can be checked rather than trusted.
- It did not fill in the gaps. Ask about something the document does not cover and it says so, and names the document it searched.
What ships
Three agent actions, ready to add to any Agentforce agent, plus one permission set. No custom objects, no scheduled jobs, no configuration screens, no data model to learn.
| Action | What it does |
|---|---|
| ZeroExport Read Attached Document | Answers a question from a document attached to the record |
| ZeroExport List Attached Documents | Lists what is attached, and what can be read |
| ZeroExport Ask Agent Help Document | Answers questions about the product itself, grounded in its own docs |
What it reads
| Format | Supported | Notes |
|---|---|---|
Word (.docx) | Yes | Headings, paragraphs, tables and lists all preserved |
PDF (.pdf) | Yes, from v0.3 | Text-based PDFs, read inside the Agentforce chat. Scanned or image-only PDFs contain no text to read |
| Anything else | No | The agent says so rather than guessing |
Limits worth knowing. Documents up to roughly 1.5 MB of internal markup, which is several hundred pages of ordinary text. Scanned or image-only documents contain no text to read, and the agent will say so rather than inventing content. Deleted tracked changes and field codes are ignored, so the agent reads the document as a person sees it.
File size is not text size. A 1.77 MB Word file we tested turned out to be 99% embedded fonts; the actual text came to 962 characters. Size on disk tells you very little about whether a document will read.
Security
- Your document is never copied out of Salesforce. Word files are opened and parsed in Apex, inside your org, using the platform's own ZIP and XML support. PDFs are processed locally in your browser by a Lightning Web Component inside the Agentforce chat, and no data is sent to any external system. No middleware touches either. No copy is stored, indexed, or retained.
- The extracted text is passed to Agentforce for a single answer, the same path as any other Agentforce prompt, through the Einstein Trust Layer. Nothing is written back, and nothing persists after the answer.
- Access follows your existing sharing rules. The agent reads as the user who asked. If someone cannot open a file themselves, the agent cannot open it for them.
- No callouts and no external endpoints. The package makes no outbound connections of any kind.
There is no ingestion step and no vector index, so nothing is copied into a second system and Data Cloud is not required.
How it works
- The agent resolves which document you mean. One attached file is used automatically, several prompt a question, and a named file is matched.
- For Word, Apex opens the
.docxwith the platform's native ZIP support and walks its XML. The document becomes clean Markdown:#headings, real tables, real lists. - For PDF, a Lightning Web Component reads the file locally in your browser and transforms it into structured JSON. This happens inside the Agentforce chat itself: there is no extra screen to open and no component to place on a page.
- That content is passed to Agentforce and grounds a single answer, cited by section.
A one-page Word document costs about 240 tokens of agent context and 23 milliseconds of processing. There is nothing to provision, schedule, or keep warm.
Next steps
Roadmap
- Larger documents, via chunked retrieval beyond the current single-transaction limit.
- More attachment sources, beyond files linked directly to a record.