docs.sheetjs.com/docz/docs/02-getting-started/02-examples/06-loader.md
2024-10-07 17:41:19 -04:00

26 KiB

title pagination_prev pagination_next sidebar_position
Loader Tutorial getting-started/installation/index getting-started/roadmap 6

import current from '/version.js'; import Tabs from '@theme/Tabs'; import TabItem from '@theme/TabItem'; import CodeBlock from '@theme/CodeBlock';

Many existing systems and platforms include support for loading data from CSV files. Many users prefer to work in spreadsheet software and multi-sheet file formats including XLSX. SheetJS libraries help bridge the gap by translating complex workbooks to simple CSV data.

The goal of this example is to load spreadsheet data into a vector store and use a large language model to generate queries based on English language input. The existing tooling supports CSV but does not support real spreadsheets.

In "SheetJS Conversion", we will use SheetJS libraries to generate CSV files for the LangChain CSV loader. These conversions can be run in a preprocessing step without disrupting existing CSV workflows.

In "SheetJS Loader", we will use SheetJS libraries in a custom LoadOfSheet data loader to directly generate documents and metadata.

"SheetJS Loader Demo" is a complete demo that uses the SheetJS Loader to answer questions based on data from a XLS workbook.

:::note Tested Deployments

This demo was tested in the following configurations:

Platform Architecture Date
NVIDIA RTX 4090 (24 GB VRAM) + i9-10910 (128 GB RAM) win10-x64 2024-08-31
NVIDIA RTX 4080 SUPER (16 GB VRAM) + i9-10910 (128 GB RAM) win10-x64 2024-08-09
AMD RX 7900 XTX (24 GB VRAM) + Ryzen Z1 Extreme (16 GB RAM) win11-x64 2024-09-21
AMD RX 6800 XT (16 GB VRAM) + Ryzen Z1 Extreme (16 GB RAM) win11-x64 2024-10-07
Apple M2 Max 12-Core CPU + 30-Core GPU (32 GB unified memory) darwin-arm 2024-07-15

SheetJS users have verified this demo in other configurations:

Other tested configurations (click to show)
Platform Architecture Demo
NVIDIA RTX 4070 Ti (12 GB VRAM) + Ryzen 7 5800x (64 GB RAM) win11-x64 LangChainJS
NVIDIA RTX 4060 (8 GB VRAM) + Ryzen 7 5700g (32 GB RAM) win11-x64 LangChainJS
NVIDIA RTX 3090 (24 GB VRAM) + Ryzen 9 3900XT (128 GB RAM) win11-x64 LangChainJS
NVIDIA RTX 3080 (12 GB VRAM) + Ryzen 7 5800X (32 GB RAM) win11-x64 LangChainJS
NVIDIA RTX 3070 (8 GB VRAM) + Ryzen Z1 Extreme (16 GB RAM) win11-x64 LangChainJS
NVIDIA RTX 3060 (12 GB VRAM) + i5-11400 (32 GB RAM) win10-x64 LangChainJS
NVIDIA RTX 2080 (12 GB VRAM) + i7-9700K (16 GB RAM) win10-x64 LangChainJS
NVIDIA RTX 2060 (6 GB VRAM) + Ryzen 5 3600 (32 GB RAM) win10-x64 LangChainJS
NVIDIA GTX 1080 (8 GB VRAM) + Ryzen 7 5800x (64 GB RAM) win10-x64 LangChainJS
NVIDIA GTX 1070 (8 GB VRAM) + Ryzen 7 7700x (32 GB RAM) win11-x64 LangChainJS

Special thanks to:

:::

CSV Loader

:::note pass

This explanation was verified against LangChain 0.2.

:::

Document loaders generate data objects ("documents") and associated metadata from data sources.

LangChain offers a CSVLoader1 component for loading CSV data from a file:

import { CSVLoader } from "@langchain/community/document_loaders/fs/csv";

const loader = new CSVLoader("pres.csv");
const docs = await loader.load();

console.log(docs);

The CSV loader uses the first row to determine column headers and generates one document per data row. For example, the following CSV holds Presidential data:

Name,Index
Bill Clinton,42
GeorgeW Bush,43
Barack Obama,44
Donald Trump,45
Joseph Biden,46

Each data row is translated to a document whose content is a list of attributes and values. For example, the third data row is shown below:

CSV RowDocument Content
Name,Index
Barack Obama,44
Name: Barack Obama
Index: 44

The LangChain CSV loader will include source metadata in the document:

Document {
  pageContent: 'Name: Barack Obama\nIndex: 44',
  metadata: { source: 'pres.csv', line: 3 }
}

SheetJS Conversion

The SheetJS NodeJS module can be imported in NodeJS scripts that use LangChain and other JavaScript libraries.

A simple pre-processing step can convert workbooks to CSV files that can be processed by the existing CSV tooling:

flowchart LR
  file[(Workbook\nXLSX/XLS)]
  subgraph SheetJS Structures
    wb(((SheetJS\nWorkbook)))
    ws((SheetJS\nWorksheet))
  end
  csv(CSV\nstring)
  docs[[Documents\nArray]]
  file --> |readFile\n\n| wb
  wb --> |wb.Sheets\nselect sheet| ws
  ws --> |sheet_to_csv\n\n| csv
  csv --> |CSVLoader\n\n| docs
  linkStyle 0,1,2 color:blue,stroke:blue;

Parsing files from the filesystem

The SheetJS readFile method2 can read workbook files. The method accepts a path and returns a workbook object that conforms to the SheetJS data model3.

/* Load SheetJS Libraries */
import { readFile, set_fs } from 'xlsx';

/* Load 'fs' for readFile support */
import * as fs from 'fs';
set_fs(fs);

/* Parse `pres.xlsx` */
// highlight-next-line
const wb = readFile("pres.xlsx");

Inspecting SheetJS workbook and worksheet objects

Workbook objects represent multi-sheet workbook files. They store individual worksheet objects and other metadata.

Relevant to this discussion, the workbook object uses the following keys4:

  • SheetNames is an array of worksheet names
  • Sheets is an object whose keys are sheet names and whose values are sheet objects.

SheetNames[0] is the first worksheet name, so the following snippet will pull the first worksheet from the workbook:

const first_ws = wb.Sheets[wb.SheetNames[0]];

Exporting SheetJS worksheets to CSV

Each worksheet in the workbook can be written to CSV text using the SheetJS sheet_to_csv5 method. The method accepts a SheetJS worksheet object and returns a string.

const csv = utils.sheet_to_csv(first_ws);

Complete Script

For example, the following NodeJS script reads pres.xlsx and displays CSV rows from the first worksheet:

/* Load SheetJS Libraries */
import { readFile, set_fs, utils } from 'xlsx';

/* Load 'fs' for readFile support */
import * as fs from 'fs';
set_fs(fs);

/* Parse `pres.xlsx` */
const wb = readFile("pres.xlsx");

/* Print CSV rows from first worksheet */
const first_ws = wb.Sheets[wb.SheetNames[0]];
const csv = utils.sheet_to_csv(first_ws);
console.log(csv);

:::note pass

A number of demos cover spiritually similar workflows:

  • Stata, MATLAB and Maple support XLSX data import. The SheetJS integrations generate clean XLSX workbooks from user-supplied spreadsheets.

  • TensorFlow.js, Pandas and Mathematica support CSV data import. The SheetJS integrations generate clean CSVs and use built-in CSV processors.

  • The "Command-Line Tools" demo covers techniques for making standalone command-line tools for file conversion.

:::

Single Worksheet

For a single worksheet, a SheetJS pre-processing step can write the CSV rows to file and the CSVLoader can load the newly written file.

Code example (click to hide)
import { CSVLoader } from "@langchain/community/document_loaders/fs/csv";
import { readFile, set_fs, utils } from 'xlsx';

/* Load 'fs' for readFile support */
import * as fs from 'fs';
set_fs(fs);

/* Parse `pres.xlsx`` */
const wb = readFile("pres.xlsx");

/* Generate CSV and write to `pres.xlsx.csv` */
const first_ws = wb.Sheets[wb.SheetNames[0]];
const csv = utils.sheet_to_csv(first_ws);
fs.writeFileSync("pres.xlsx.csv", csv);

/* Create documents with CSVLoader */
const loader = new CSVLoader("pres.xlsx.csv");
const docs = await loader.load();

console.log(docs);
// ...

Workbook

A workbook is a collection of worksheets. Each worksheet can be exported to a separate CSV. If the CSVs are written to a subfolder, a DirectoryLoader6 can process the files in one step.

Code example (click to hide)

In this example, the script creates a subfolder named csv. Each worksheet in the workbook will be processed and the generated CSV will be stored to numbered files. The first worksheet will be stored to csv/0.csv.

import { CSVLoader } from "@langchain/community/document_loaders/fs/csv";
import { DirectoryLoader } from "langchain/document_loaders/fs/directory";
import { readFile, set_fs, utils } from 'xlsx';

/* Load 'fs' for readFile support */
import * as fs from 'fs';
set_fs(fs);

/* Parse `pres.xlsx`` */
const wb = readFile("pres.xlsx");

/* Create a folder `csv` */
try { fs.mkdirSync("csv"); } catch(e) {}

/* Generate CSV data for each worksheet */
wb.SheetNames.forEach((name, idx) => {
  const ws = wb.Sheets[name];
  const csv = utils.sheet_to_csv(ws);
  fs.writeFileSync(`csv/${idx}.csv`, csv);
});

/* Create documents with DirectoryLoader */
const loader = new DirectoryLoader("csv", {
  ".csv": (path) => new CSVLoader(path)
});
const docs = await loader.load();

console.log(docs);
// ...

SheetJS Loader

The CSVLoader that ships with LangChain does not add any Document metadata and does not generate any attributes. A custom loader can work around limitations in the CSV tooling and potentially include metadata that has no CSV equivalent.

flowchart LR
  file[(Workbook\nXLSX/XLS)]
  subgraph SheetJS Structures
    wb(((SheetJS\nWorkbook)))
    ws((SheetJS\nWorksheet))
  end
  aoo[(Array of\nObjects)]
  docs[[Documents\nArray]]
  file --> |readFile\n\n| wb
  wb --> |wb.Sheets\nEach worksheet| ws
  ws --> |sheet_to_json\n\n| aoo
  aoo --> |new Document\nEach Row| docs
  linkStyle 0,1,2 color:blue,stroke:blue;

The demo LoadOfSheet loader will generate one Document per data row across all worksheets. It will also attempt to build metadata and attributes for use in self-querying retrievers.

/* read and parse `data.xlsb` */
const loader = new LoadOfSheet("./data.xlsb");

/* generate documents */
const docs = await loader.load();

/* synthesized attributes for the SelfQueryRetriever */
const attributes = loader.attributes;
Sample SheetJS Loader (click to show)

This example loader pulls data from each worksheet. It assumes each worksheet includes one header row and a number of data rows.

import { Document } from "@langchain/core/documents";
import { BufferLoader } from "langchain/document_loaders/fs/buffer";
import { read, utils } from "xlsx";

/**
 * Document loader that uses SheetJS to load documents.
 *
 * Each worksheet is parsed into an array of row objects using the SheetJS
 * `sheet_to_json` method and projected to a `Document`. Metadata includes
 * original sheet name, row data, and row index
 */
export default class LoadOfSheet extends BufferLoader {
  /** @type {import("langchain/chains/query_constructor").AttributeInfo[]}  */
  attributes = [];

  /**
   * Document loader that uses SheetJS to load documents.
   *
   * @param {string|Blob} filePathOrBlob Source Data
   */
  constructor(filePathOrBlob) {
    super(filePathOrBlob);
    this.attributes = [];
  }

  /**
   * Parse document
   *
   * NOTE: column labels in multiple sheets are not disambiguated!
   *
   * @param {Buffer} raw Raw data Buffer
   * @param {Document["metadata"]} metadata Document metadata
   * @returns {Promise<Document[]>} Array of Documents
   */
  async parse(raw, metadata) {
    /** @type {Document[]} */
    const result = [];

    this.attributes = [
      { name: "worksheet", description: "Sheet or Worksheet Name", type: "string" },
      { name: "rowNum", description: "Row index", type: "number" }
    ];

    const wb = read(raw, {type: "buffer", WTF:1});
    for(let name of wb.SheetNames) {
      const fields = {};
      const ws = wb.Sheets[name];
      if(!ws) return;

      const aoo = utils.sheet_to_json(ws);
      aoo.forEach((row, idx) => {
        result.push({
          pageContent: "Row " + (idx + 1) + " has the following content: \n" + Object.entries(row).map(kv => `- ${kv[0]}: ${kv[1]}`).join("\n") + "\n",
          metadata: {
            worksheet: name,
            rowNum: row["__rowNum__"],
            ...metadata,
            ...row
          }
        });
        Object.entries(row).forEach(([k,v]) => { if(v != null) (fields[k] || (fields[k] = {}))[v instanceof Date ? "date" : typeof v] = true } );
      });
      Object.entries(fields).forEach(([k,v]) => this.attributes.push({
        name: k, description: k, type: Object.keys(v).join(" or ")
      }));
    }

    return result;
  }
};

From Text to Binary

Many libraries and platforms offer generic "text" loaders that process files assuming the UTF8 encoding. This corrupts many spreadsheet formats including XLSX, XLSB, XLSM and XLS.

:::note pass

This issue affects many JavaScript tools. Various demos cover workarounds:

:::

The CSVLoader extends a special TextLoader that forces UTF8 text parsing.

There is a separate BufferLoader class, used by the PDF loader, that passes the raw data using NodeJS Buffer objects.

BinaryText
export class PDFLoader extends BufferLoader {
  // ...
  public async parse(
    raw: Buffer,
    metadata: Document["metadata"]
  ): Promise<Document[]> {
    // ...
  }
  // ...
}
export class CSVLoader extends TextLoader {
  // ...
  protected async parse(
    raw: string

  ): Promise<string[]> {
    // ...
  }
  // ...
}

NodeJS Buffers

The SheetJS read method supports NodeJS Buffer objects directly7:

import { BufferLoader } from "langchain/document_loaders/fs/buffer";
import { read, utils } from "xlsx";

export default class LoadOfSheet extends BufferLoader {
  // ...
  async parse(raw, metadata) {
    // highlight-next-line
    const wb = read(raw, {type: "buffer"});
    // At this point, `wb` is a SheetJS workbook object
    // ...
  }
}

The read method returns a SheetJS workbook object4.

Generating Content

The SheetJS sheet_to_json method8 returns an array of data objects whose keys are drawn from the first row of the worksheet.

SpreadsheetArray of Objects

pres.xlsx data

[
  { Name: "Bill Clinton", Index: 42 },
  { Name: "GeorgeW Bush", Index: 43 },
  { Name: "Barack Obama", Index: 44 },
  { Name: "Donald Trump", Index: 45 },
  { Name: "Joseph Biden", Index: 46 }
]

The original CSVLoader wrote one row for each key-value pair. This text can be generated by looping over the keys and values of the data row object. The Object.entries helper function simplifies the conversion:

function make_csvloader_doc_from_row_object(row) {
  return Object.entries(row).map(([k,v]) => `${k}: ${v}`).join("\n");
}

Generating Documents

The loader must generate row objects for each worksheet in the workbook.

In the SheetJS data model, the workbook object has two relevant fields:

  • SheetNames is an array of sheet names
  • Sheets is an object whose keys are sheet names and values are sheet objects.

A for..of loop can iterate across the worksheets:

    const wb = read(raw, {type: "buffer", WTF:1});
    for(let name of wb.SheetNames) {
      const ws = wb.Sheets[name];
      const aoa = utils.sheet_to_json(ws);
      // at this point, `aoa` is an array of objects
    }

This simplified parse function uses the snippet from the previous section:

  async parse(raw, metadata) {
    /* array to hold generated documents */
    const result = [];

    /* read workbook */
    const wb = read(raw, {type: "buffer", WTF:1});

    /* loop over worksheets */
    for(let name of wb.SheetNames) {
      const ws = wb.Sheets[name];
      const aoa = utils.sheet_to_json(ws);

      /* loop over data rows */
      aoa.forEach((row, idx) => {
        /* generate a new document and add to the result array */
        result.push({
          pageContent: Object.entries(row).map(([k,v]) => `${k}: ${v}`).join("\n")
        });
      });
    }

    return result;
  }

Metadata and Attributes

It is strongly recommended to generate additional metadata and attributes for self-query retrieval applications.

Implementation Details (click to show)

Metadata

Metadata is attached to each document object. The following example appends the raw row data to the document metadata:

        /* generate a new document and add to the result array */
        result.push({
          pageContent: Object.entries(row).map(([k,v]) => `${k}: ${v}`).join("\n"),
          metadata: {
            worksheet: name, // name of the worksheet
            rowNum: idx, // data row index
            ...row // raw row data
          }
        });

Attributes

Each attribute object specifies three properties:

  • name corresponds to the field in the document metadata
  • description is a description of the field
  • type is a description of the data type.

While looping through data rows, a simple type check can keep track of the data type for each column:

    for(let name of wb.SheetNames) {
      /* track column types */
      const fields = {};
      // ...

      aoo.forEach((row, idx) => {
        result.push({/* ... */});
        /* Check each property */
        Object.entries(row).forEach(([k,v]) => {
          /* Update fields entry to reflect the new data point */
          if(v != null) (fields[k] || (fields[k] = {}))[v instanceof Date ? "date" : typeof v] = true
        });
      });
      // ...
    }

Attributes can be generated after writing the worksheet data. Storing attributes in a loader property will make it accessible to scripts that use the loader.

export default class LoadOfSheet extends BufferLoader {
  // highlight-next-line
  attributes = [];
  // ...

  async parse(raw, metadata) {
    // Add the worksheet name and row index attributes
    // highlight-start
    this.attributes = [
      { name: "worksheet", description: "Sheet or Worksheet Name", type: "string" },
      { name: "rowNum", description: "Row index", type: "number" }
    ];
    // highlight-end
    const wb = read(raw, {type: "buffer", WTF:1});
    for(let name of wb.SheetNames) {
      // highlight-next-line
      const fields = {};
      // ...
      const aoo = utils.sheet_to_json(ws);
      aoo.forEach((row, idx) => {
        result.push({/* ... */});
        /* Check each property */
        Object.entries(row).forEach(([k,v]) => {
          /* Update fields entry to reflect the new data point */
          if(v != null) (fields[k] || (fields[k] = {}))[v instanceof Date ? "date" : typeof v] = true
        });
      });
      /* Add one attribute per metadata field */
      // highlight-start
      Object.entries(fields).forEach(([k,v]) => this.attributes.push({
        name: k, description: k,
        /* { number: true, string: true } -> "number or string" */
        type: Object.keys(v).join(" or ")
      }));
      // highlight-end
    }
    // ...
  }

SheetJS Loader Demo

The demo performs the query "Which rows have over 40 miles per gallon?" against a sample cars dataset and displays the results.

:::note pass

SheetJS team members have tested this demo on Windows 10 and Windows 11 using PowerShell and Ollama for Windows.

SheetJS users have also tested this demo within Windows Subsystem for Linux.

:::

:::caution pass

This demo was tested using the ChatQA-1.5 model9 in Ollama.

The tested model used up to 9.2GB VRAM. It is strongly recommended to run the demo on a newer Apple Silicon Mac or a PC with an Nvidia GPU with at least 12GB VRAM. SheetJS users have tested the demo on systems with as little as 6GB VRAM.

:::

  1. Install pre-requisites:

:::note pass

Ollama should be installed on the same platform as NodeJS. If NodeJS is run within WSL, Ollama should also be installed within WSL.

:::

After installing dependencies, start a new terminal session.

  1. Create a new project:
mkdir sheetjs-loader
cd sheetjs-loader
npm init -y
  1. Download the demo scripts:
curl -LO https://docs.sheetjs.com/loadofsheet/query.mjs
curl -LO https://docs.sheetjs.com/loadofsheet/loadofsheet.mjs

:::note pass

In PowerShell, the command may fail with a parameter error:

Invoke-WebRequest : A parameter cannot be found that matches parameter name 'LO'.

curl.exe must be invoked directly:

curl.exe -LO https://docs.sheetjs.com/loadofsheet/query.mjs
curl.exe -LO https://docs.sheetjs.com/loadofsheet/loadofsheet.mjs

:::

  1. Install the SheetJS NodeJS module:

{\ npm i --save https://sheet.lol/balls/xlsx-${current}.tgz}

  1. Install dependencies:
npm i --save @langchain/community@0.2.18 @langchain/core@0.2.15 langchain@0.2.9 peggy@4.0.3 --force

:::note pass

When this demo was last tested, there were error messages relating to dependency and peer dependency versions. The --force flag was required.

:::

  1. Download the cars dataset:
curl -LO https://docs.sheetjs.com/cd.xls

:::note pass

In PowerShell, the command may fail with a parameter error:

Invoke-WebRequest : A parameter cannot be found that matches parameter name 'LO'.

curl.exe must be invoked directly:

curl.exe -LO https://docs.sheetjs.com/cd.xls

:::

  1. Install the llama3-chatqa:8b-v1.5-q8_0 model using Ollama:
ollama pull llama3-chatqa:8b-v1.5-q8_0
  1. Run the demo script
node query.mjs

The demo performs the query "Which rows have over 40 miles per gallon?". It will print the following nine results:

{ Name: 'volkswagen rabbit custom diesel', MPG: 43.1 }
{ Name: 'vw rabbit c (diesel)', MPG: 44.3 }
{ Name: 'renault lecar deluxe', MPG: 40.9 }
{ Name: 'honda civic 1500 gl', MPG: 44.6 }
{ Name: 'datsun 210', MPG: 40.8 }
{ Name: 'vw pickup', MPG: 44 }
{ Name: 'mazda glc', MPG: 46.6 }
{ Name: 'vw dasher (diesel)', MPG: 43.4 }
{ Name: 'vw rabbit', MPG: 41.5 }

:::caution pass

Some SheetJS users with older GPUs have reported errors.

If the command fails, please try running the script a second time.

:::

To find the expected results:

  • Open the cd.xls spreadsheet in Excel
  • Select Home > Sort & Filter > Filter in the Ribbon
  • Select the filter option for column B (Miles_per_Gallon)
  • In the popup, select "Greater Than" in the Filter dropdown and type 40

The filtered results should match the following screenshot:

Expected Results

:::tip pass

The SheetJS model exposes formulae and other features.

SheetJS Pro builds expose cell styling, images, charts, tables, and other features.

:::