PDF Table Data Conversion

Turn tables locked inside a PDF into a clean, nicely formatted spreadsheet the user can open in Excel or feed to another tool. The user is asking for the data as a file, not a prose answer — so your job is to locate the source PDF, get its full content (not just retrieved chunks), extract the relevant table(s), and hand back a formatted workbook.

When this applies (and when it does not)

• Applies — the user wants the tabular data itself as a file: "pull out / extract / export / convert / get me a spreadsheet of …", "put the rebate schedule in Excel", "give me the line items from this contract".
• Applies to follow-ups too — even if you already answered or summarized the data in the chat, a subsequent "give me that as a file / spreadsheet / Excel" is still an extraction request. Invoke this skill and rebuild the file from the source document — do not assemble a spreadsheet from the text already in the conversation, which may be partial, truncated, or reformatted.
• Does not apply — the user is asking a question to be answered in the conversation: "what are the rebates for Contoso?", "how much is the Q3 discount?". Answer those inline from knowledge within reason; do not run an extraction or produce a file.

If a request is ambiguous (e.g. "show me the rebates"), prefer a short inline answer and offer to export it as a file if they want the full table.

Instructions

1. Identify the target document. Let the user ask naturally — they need not name an exact file. Use the cues in their request (customer/contract name, topic, document title, or a file they referenced) to search knowledge for the best-matching PDF. If one document is a clear match, proceed with it. Only when the match is genuinely ambiguous — several plausible documents, or none obvious — ask the user a brief question to confirm which one they mean.
2. Get the FULL document, not chunks. Retrieved knowledge chunks are enough to answer a question but not to extract a whole table — rows are routinely split across chunks — so work from the complete PDF in the agent's container. A SharePoint knowledge source handles this naturally: let it search and pull the full file down. If the maker collects files another way (direct upload, a connector, a different store), use whatever mechanism is available to get the full PDF locally. The rest of these steps are the same once the file is local.
3. Extract the table(s) — script first. Run the bundled scripts/extract_tables.py against the downloaded PDF to pull tabular data out deterministically (see Bundled files). This is the default path because it lifts the grid verbatim — no dropped or "tidied" rows — which matters most for long schedules. It writes a formatted .xlsx workbook by default, with one tab per detected table (so multiple tables and page splits stay cleanly separated). Narrow to the relevant table when the user named a specific one (e.g. rebates, pricing, line items) using the --contains filter; otherwise extract all detected tables.
4. Fall back to reading it yourself when the script can't cope. pdfplumber relies on a text layer and clean rules, so it under-performs on scanned / image-only PDFs (no text to extract) and borderless or merged-cell tables (misaligned output). When the script returns nothing, or the output is clearly garbled/misaligned versus the PDF, read the table directly from the document and build the workbook yourself following the cleaning rules below. Use your judgement to pick the right table when a keyword filter is too blunt.
5. Clean the output. Ensure a single header row per table, trim stray whitespace, drop fully empty rows (but keep empty columns — they preserve the table's structure and alignment), and keep numbers/currency/dates as they appear in the source (do not invent, reformat, or "correct" values). If cells were merged or a header spans multiple rows, flatten to one clear header row. Put each distinct table on its own sheet/tab; if one logical table is split across pages, you may merge the parts into a single sheet.
6. Respond with a recap, then the file. Open your reply with a short, natural summary of what the user asked for and what you pulled — restate the request in your own words (e.g. "Here's the full rebate schedule from the Contoso contract you asked for") — then attach the full extracted data as a formatted .xlsx workbook. The complete data lives in the file, not the chat: don't paste the whole table inline or truncate it. When multiple tables were extracted, give each its own clearly named tab and briefly say what each contains. Then offer a .csv as an alternative ("Let me know if this works, or if you'd prefer a CSV instead") — produce it with --format csv (or both) if they say yes, one CSV per table.
7. Confirm and flag gaps. Give a one-line summary (which document, which table(s)/tab(s), rows × columns) and note whether the script or a manual read produced it. Call out anything uncertain — a table that spanned pages, cells that failed to parse, or a table the tool could not detect — so the user can verify rather than trust silently.

Bundled files

The attached .zip includes:

• scripts/extract_tables.py — extracts tables from a local PDF using pdfplumber (available natively in the agent container). Run it as python scripts/extract_tables.py <document.pdf> --out-dir out. By default it writes a formatted .xlsx workbook — one tab per table, with a bold frozen header row, an autofilter, and sized columns. Useful flags:
  • --format xlsx|csv|both — output format (default xlsx; csv writes one CSV per table; both writes the workbook and the CSVs),
  • --pages 2-5 (or --pages 3) to limit to specific pages,
  • --contains rebate to keep only tables whose text contains a keyword,
  • --out-dir out to choose where the files are written. It prints a JSON summary of everything produced (workbook path, per-table sheet names, any CSV paths, source page, row/column counts) so you can report back accurately. If it produces no tables or clearly mangles them (scanned or borderless PDFs), fall back to reading the tables yourself per step 4.

Guardrails

• Never fabricate rows, values, or headers. If a cell is unreadable, leave it empty and flag it rather than guessing.
• Do not answer extraction requests from retrieved chunks alone — always work from the full downloaded document so tables aren't truncated.
• Never build the file from the chat. When the user asks for a file after you've already answered or summarized in conversation, still run this skill and extract from the source document — do not hand-assemble a spreadsheet from the text in the chat, which may be partial or reformatted.
• Do not turn a plain question into a file dump; only produce a file when the user actually wants the data as a file.
• Do not expand scope beyond what was asked (don't export every table when the user asked only for the rebates).

Tone

Precise and practical. Prefer surfacing uncertainty over confidently returning a table that may be incomplete.

Run this — do not improvise

This capability's deterministic implementation is a RAPP single-file agent, linked beside this file as pdf_table_data_conversion_agent.py and embedded as the fenced Python below (sha256 815527a7c5b5db5a…; a byte-exact copy is also vaulted in the capsule comment at the end of this file). On a host with sandbox execution, run the linked file directly — if it is missing, write the fence contents verbatim to pdf_table_data_conversion_agent.py first:

python3 pdf_table_data_conversion_agent.py '{"key": "value"}'      # arguments as one JSON object
echo '{"key": "value"}' | python3 pdf_table_data_conversion_agent.py   # or on stdin
python3 pdf_table_data_conversion_agent.py --tool                      # emit the JSON tool contract

Treat stdout as a tool result. If it reports missing or unresolved inputs, stop and collect them. If it returns steps, execute those steps in order exactly as returned; if it returns instructions, follow them with the supplied inputs. Otherwise use the result verbatim. Do not invent behavior beyond that output. On a host without code execution, treat the Parameters schema and the code below as the exact specification and never paraphrase a step. Never edit inside the generated markers; a converter-equipped host can instead restore the original file checksum-verified with the installed rapp-agent-converter/scripts/toast.py convert SKILL.md --to agent.

"""PdfTableDataConversion -- Use this skill whenever the user wants tabular data from a PDF document AS A FILE — to extract, pull out, export, convert, or "put into a spreadsheet/Excel/CSV" (e.g. "pull out the rebates from the Contoso contract", "extract this pricing table", "get this into a spreadsheet"). This INCLUDES follow-up requests after you have already answered or summarized the data in chat — e.g. "now give me that as a file/spreadsheet", "export what you just showed me", "download that as Excel". In those follow-ups you MUST still invoke this skill and extract from the source document; never hand-build a spreadsheet from the chat summary. Do NOT use this skill for answer-in-line questions such as "what are the rebates in the Contoso contract?" — those are grounded Q&A.

Generated by the rapp skill from pdf-table-data-conversion. The RCI capsule at the bottom of this file carries the full original; `toast.py convert` restores it byte-exact."""

import json
import re
import sys

try:
    from agents.basic_agent import BasicAgent
except ImportError:  # running OUTSIDE a brainstem -- stay executable anyway.
    class BasicAgent:  # noqa: D101 - minimal stand-in, same contract
        def __init__(self, name=None, metadata=None):
            if name:
                self.name = name
            if metadata:
                self.metadata = metadata

        def perform(self, **kwargs):
            return "Not implemented."

        def system_context(self):
            return None

        def to_tool(self):
            return {"type": "function", "function": {
                "name": self.name,
                "description": self.metadata.get("description", ""),
                "parameters": self.metadata.get("parameters", {})}}

# The procedural layer, verbatim from the source capability.
INSTRUCTIONS = 'Turn tables locked inside a PDF into a clean, nicely formatted spreadsheet the\nuser can open in Excel or feed to another tool. The user is asking for the *data\nas a file*, not a prose answer — so your job is to locate the source PDF, get its\n**full** content (not just retrieved chunks), extract the relevant table(s), and\nhand back a formatted workbook.\n\n## When this applies (and when it does not)\n- **Applies** — the user wants the tabular data itself as a file: "pull out /\n  extract / export / convert / get me a spreadsheet of …", "put the rebate\n  schedule in Excel", "give me the line items from this contract".\n- **Applies to follow-ups too** — even if you already answered or summarized the\n  data in the chat, a subsequent "give me that as a file / spreadsheet / Excel"\n  is still an extraction request. Invoke this skill and rebuild the file from the\n  **source document** — do not assemble a spreadsheet from the text already in the\n  conversation, which may be partial, truncated, or reformatted.\n- **Does not apply** — the user is asking a question to be answered in the\n  conversation: "what are the rebates for Contoso?", "how much is the Q3\n  discount?". Answer those inline from knowledge within reason; do **not** run an\n  extraction or produce a file.\n\nIf a request is ambiguous (e.g. "show me the rebates"), prefer a short inline\nanswer and offer to export it as a file if they want the full table.\n\n## Instructions\n1. **Identify the target document.** Let the user ask naturally — they need not\n   name an exact file. Use the cues in their request (customer/contract name,\n   topic, document title, or a file they referenced) to search knowledge for the\n   best-matching PDF. If one document is a clear match, proceed with it. Only when\n   the match is genuinely ambiguous — several plausible documents, or none obvious\n   — ask the user a brief question to confirm which one they mean.\n2. **Get the FULL document, not chunks.** Retrieved knowledge chunks are enough\n   to *answer* a question but not to *extract a whole table* — rows are routinely\n   split across chunks — so work from the complete PDF in the agent's container.\n   A SharePoint knowledge source handles this naturally: let it search and pull\n   the full file down. If the maker collects files another way (direct upload, a\n   connector, a different store), use whatever mechanism is available to get the\n   full PDF locally. The rest of these steps are the same once the file is local.\n3. **Extract the table(s) — script first.** Run the bundled\n   `scripts/extract_tables.py` against the downloaded PDF to pull tabular data out\n   deterministically (see *Bundled files*). This is the default path because it\n   lifts the grid verbatim — no dropped or "tidied" rows — which matters most for\n   long schedules. It writes a formatted `.xlsx` workbook by default, with one tab\n   per detected table (so multiple tables and page splits stay cleanly separated).\n   Narrow to the relevant table when the user named a specific one (e.g. rebates,\n   pricing, line items) using the `--contains` filter; otherwise extract all\n   detected tables.\n4. **Fall back to reading it yourself when the script can't cope.** `pdfplumber`\n   relies on a text layer and clean rules, so it under-performs on **scanned /\n   image-only PDFs** (no text to extract) and **borderless or merged-cell tables**\n   (misaligned output). When the script returns nothing, or the output is clearly\n   garbled/misaligned versus the PDF, read the table directly from the document\n   and build the workbook yourself following the cleaning rules below. Use your\n   judgement to pick the right table when a keyword filter is too blunt.\n5. **Clean the output.** Ensure a single header row per table, trim stray\n   whitespace, drop fully empty rows (but keep empty columns — they preserve the\n   table's structure and alignment), and keep numbers/currency/dates as they\n   appear in the source (do not invent, reformat, or "correct" values). If cells\n   were merged or a header spans multiple rows, flatten to one clear header row.\n   Put each distinct table on its own sheet/tab; if one logical table is split\n   across pages, you may merge the parts into a single sheet.\n6. **Respond with a recap, then the file.** Open your reply with a short,\n   natural summary of what the user asked for and what you pulled — restate the\n   request in your own words (e.g. "Here's the full rebate schedule from the\n   Contoso contract you asked for") — then attach the **full** extracted data as a\n   formatted `.xlsx` workbook. The complete data lives in the file, not the chat:\n   don't paste the whole table inline or truncate it. When multiple tables were\n   extracted, give each its own clearly named tab and briefly say what each\n   contains. Then **offer a `.csv` as an alternative** ("Let me know if this works, or if you'd prefer a CSV instead") — produce it with `--format csv` (or `both`) if they say yes, one\n   CSV per table.\n7. **Confirm and flag gaps.** Give a one-line summary (which document, which\n   table(s)/tab(s), rows × columns) and note whether the script or a manual read\n   produced it. Call out anything uncertain — a table that spanned pages, cells\n   that failed to parse, or a table the tool could not detect — so the user can\n   verify rather than trust silently.\n\n## Bundled files\nThe attached `.zip` includes:\n- `scripts/extract_tables.py` — extracts tables from a local PDF using\n  `pdfplumber` (available natively in the agent container). Run it as\n  `python scripts/extract_tables.py <document.pdf> --out-dir out`. By default it\n  writes a formatted `.xlsx` workbook — one tab per table, with a bold frozen\n  header row, an autofilter, and sized columns. Useful flags:\n  - `--format xlsx|csv|both` — output format (default `xlsx`; `csv` writes one\n    CSV per table; `both` writes the workbook and the CSVs),\n  - `--pages 2-5` (or `--pages 3`) to limit to specific pages,\n  - `--contains rebate` to keep only tables whose text contains a keyword,\n  - `--out-dir out` to choose where the files are written.\n  It prints a JSON summary of everything produced (workbook path, per-table sheet\n  names, any CSV paths, source page, row/column counts) so you can report back\n  accurately. If it produces no tables or clearly\n  mangles them (scanned or borderless PDFs), fall back to reading the tables\n  yourself per step 4.\n\n## Guardrails\n- Never fabricate rows, values, or headers. If a cell is unreadable, leave it\n  empty and flag it rather than guessing.\n- Do not answer extraction requests from retrieved chunks alone — always work\n  from the full downloaded document so tables aren't truncated.\n- **Never build the file from the chat.** When the user asks for a file after\n  you've already answered or summarized in conversation, still run this skill and\n  extract from the source document — do not hand-assemble a spreadsheet from the\n  text in the chat, which may be partial or reformatted.\n- Do not turn a plain question into a file dump; only produce a file when the user\n  actually wants the data as a file.\n- Do not expand scope beyond what was asked (don't export every table when the\n  user asked only for the rebates).\n\n## Tone\nPrecise and practical. Prefer surfacing uncertainty over confidently returning a\ntable that may be incomplete.'

# Ordered commands lifted verbatim from the capability's own documentation.
STEPS = []


class PdfTableDataConversionAgent(BasicAgent):
    def __init__(self):
        self.name = 'PdfTableDataConversion'
        self.metadata = {
          "name": "PdfTableDataConversion",
          "description": "Use this skill whenever the user wants tabular data from a PDF document AS A FILE \u2014 to extract, pull out, export, convert, or \"put into a spreadsheet/Excel/CSV\" (e.g. \"pull out the rebates from the Contoso contract\", \"extract this pricing table\", \"get this into a spreadsheet\"). This INCLUDES follow-up requests after you have already answered or summarized the data in chat \u2014 e.g. \"now give me that as a file/spreadsheet\", \"export what you just showed me\", \"download that as Excel\". In those follow-ups you MUST still invoke this skill and extract from the source document; never hand-build a spreadsheet from the chat summary. Do NOT use this skill for answer-in-line questions such as \"what are the rebates in the Contoso contract?\" \u2014 those are grounded Q&A.",
          "parameters": {
            "type": "object",
            "properties": {},
            "required": []
          }
        }
        super().__init__(name=self.name, metadata=self.metadata)

    def perform(self, **kwargs):  # toaster:generated-perform
        return json.dumps({"status": "ok", "instructions": INSTRUCTIONS,
                           "inputs": kwargs,
                           "note": "Prose-only capability: follow INSTRUCTIONS "
                                   "with the given inputs."}, indent=2)

if __name__ == "__main__":
    #     echo '{"arg": "value"}' | python3 pdf_table_data_conversion_agent.py
    #     python3 pdf_table_data_conversion_agent.py '{"arg": "value"}'
    #     python3 pdf_table_data_conversion_agent.py --tool          # emit the JSON tool contract
    _a = sys.argv[1:]
    if _a and _a[0] == "--tool":
        print(json.dumps(PdfTableDataConversionAgent().to_tool(), indent=2))
    else:
        _raw = _a[0] if _a else (sys.stdin.read().strip() or "{}")
        print(PdfTableDataConversionAgent().perform(**json.loads(_raw)))

# rci-capsule:v1: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