Skip to content

Tutorial 15 - Investigate Documents with a Pipeline

Time 6 min · Difficulty ★★☆ · Surfaces Agentic Chat

Goal

Point the agent at a document it has never seen and let it dig the facts out with filesystem primitives - the Unix pipeline way (find | grep -C | wc -l), one tool call per step. You apply the Document detective preset and ask about a contract you have not pasted; the model calls requestFileUpload, you pick the file, and the agent hunts the clause down with grepFile and sliceFile, quoting it back with exact line numbers.

Steps

  1. Open Agentic Chat, click the Prompt Library (clipboard) icon, and pick the Document detective preset from the Presets group. Apply to chat loads it into the Agentic Chat Setting drawer; Apply & New Chat commits it, and the Apply preset tools dialog then asks to publish the preset's twelve tools - requestFileUpload plus the read pipeline (findFiles, searchInFiles, grepFile, sliceFile, lineCount, ...). Click Apply.

  2. Ask about a document you have not pasted - for example I have a contract file. Find the late delivery penalty clause and quote it with its exact line numbers. The model realizes it cannot read what it does not have and calls requestFileUpload - the chat opens an Upload a file dialog mid-conversation.

    Use a tool-capable model

    This preset asks the model to chain several filesystem calls, which is more than the shipped qwen3.5:4b default reliably manages: it tends to narrate the plan ("Step 1: check allowed directories") as text instead of calling the tools, and then no upload dialog ever opens. Switch to qwen3.5:9b or gemma4:e4b in Agentic Chat Setting before you start - the run below is qwen3.5:9b.

The Upload a file dialog opened by requestFileUpload, asking for the contract file The model writes the ask itself - here Please upload your contract file so we can analyze it for late delivery penalty clauses. The same human-in-the-loop seam as tool approvals: the turn pauses until you hand over a file or cancel. Excel converts to CSV in the browser before anything is stored.

  1. Pick the file. It lands under the conversation workspace at uploads/<name> and the tool returns that path to the model, which reads it and answers with the clause quoted at its exact line numbers - evidence, not paraphrase. The preset scales the tooling to the file: a small contract like this one is read whole with readTextFile; on a large document it switches to grepFile numbered=true to pin the matching line and sliceFile to pull a context window around it.

The agent's answer - the clause quoted, Line Numbers reading Lines 5-6, an Evidence Quote per line, and a numbered Pipeline Steps list of the six tool calls Every claim carries a verbatim quote and a line number - the grep -n discipline as an agent behavior - and the answer ends with the Pipeline Steps it actually ran: listAllowedDirectories() to find the roots, requestFileUpload(...) to get the file, readTextFile(...), lineCount(...), grepFile(...) to pin the match, then sliceFile(start=3, end=8) for the context window. The run meta above it reports the cost of that discipline - here 135.8s · 27,908 tokens.

  1. No upload needed when the documents are already on disk: give it a folder path instead - Search every file under ~/contracts for the word "penalty" - and the agent sweeps it with findFiles + searchInFiles, then slices context around each hit, exactly like the preset card's captured run.

What to observe

  • The upload is saved under the conversation workspace (workspace/<conversationId>/uploads/) - each chat keeps its own files, and deleting the conversation removes them.
  • The pipeline idiom maps one-to-one: findFiles = find, grepFile numbered=true = grep -n, sliceFile = sed -n 'a,bp', lineCount = wc -l, sortFile / cutFileFields = sort / cut.
  • Everything here is read-only (L3) - unlike the write tools in Tutorial 13, no approval prompt is needed to search and quote.
  • Excel (.xlsx / .xls) converts to CSV in the browser before upload; the server only ever stores text.

Why this matters

Contracts, reports, exports, logs from another machine - most real documents are files, not paste-able text. The upload dialog plus the read pipeline turns "let me summarize what I think it says" into "line 5 says exactly this", with the evidence trail to prove it. The full preset reference is at Prompt Presets → Document detective.

Was this page helpful? Ask on Discussions · Report a docs issue