How do you mail merge a word document in c#

This is quite simple by using Microsoft.Office.Interop.Word. Here is a simple step by step tutorial on how to do this.

The code to replace a mergefield with a string is like this:

public static void TextToWord(string pWordDoc, string pMergeField, string pValue)
{
    Object oMissing = System.Reflection.Missing.Value;
    Object oTrue = true;
    Object oFalse = false;
    Word.Application oWord = new Word.Application();
    Word.Document oWordDoc = new Word.Document();
    oWord.Visible = true;
    Object oTemplatePath = pWordDoc;
    oWordDoc = oWord.Documents.Add(ref oTemplatePath, ref oMissing, ref oMissing, ref oMissing);
    foreach (Word.Field myMergeField in oWordDoc.Fields)
    {
        Word.Range rngFieldCode = myMergeField.Code;
        String fieldText = rngFieldCode.Text;
        if (fieldText.StartsWith(" MERGEFIELD"))
        {
            Int32 endMerge = fieldText.IndexOf("\\");
            Int32 fieldNameLength = fieldText.Length - endMerge;
            String fieldName = fieldText.Substring(11, endMerge - 11);
            fieldName = fieldName.Trim();
            if (fieldName == pMergeField)
            {
                myMergeField.Select();
                oWord.Selection.TypeText(pValue);
            }
        }
    }
}

originally posted here and here

In case you wish to use a dictionary to replace many fields at once use the code below:

public static void TextToWord(string pWordDoc, Dictionary<string, string> pDictionaryMerge)
    {
        Object oMissing = System.Reflection.Missing.Value;
        Object oTrue = true;
        Object oFalse = false;
        Microsoft.Office.Interop.Word.Application oWord = new Microsoft.Office.Interop.Word.Application();
        Microsoft.Office.Interop.Word.Document oWordDoc = new Microsoft.Office.Interop.Word.Document();
        oWord.Visible = true;
        Object oTemplatePath = pWordDoc;
        oWordDoc = oWord.Documents.Add(ref oTemplatePath, ref oMissing, ref oMissing, ref oMissing);

        foreach (Microsoft.Office.Interop.Word.Field myMergeField in oWordDoc.Fields)
        {
            Microsoft.Office.Interop.Word.Range rngFieldCode = myMergeField.Code;
            String fieldText = rngFieldCode.Text;
            if (fieldText.StartsWith(" MERGEFIELD"))
            {
                Int32 endMerge = fieldText.IndexOf("\\");
                Int32 fieldNameLength = fieldText.Length - endMerge;
                String fieldName = fieldText.Substring(11, endMerge - 11);
                fieldName = fieldName.Trim();
                foreach (var item in pDictionaryMerge)
                {
                    if (fieldName == item.Key)
                    {
                        myMergeField.Select();
                        oWord.Selection.TypeText(item.Value);
                    }
                }
            }
        }
    }

Can't believe third party software charge thousands for interface functions with Word. I have perfectly solved this mail merge thing in my project -- no third party, no particular demands on IIS, just use OpenXML. So, add these 4 functions to your project:

public static void dotx2docx(string sourceFile, string targetFile)
    {
        MemoryStream documentStream;
        using (Stream tplStream = File.OpenRead(sourceFile))
        {
            documentStream = new MemoryStream((int)tplStream.Length);
            CopyStream(tplStream, documentStream);
            documentStream.Position = 0L;
        }

        using (WordprocessingDocument template = WordprocessingDocument.Open(documentStream, true))
        {
            template.ChangeDocumentType(DocumentFormat.OpenXml.WordprocessingDocumentType.Document);
            MainDocumentPart mainPart = template.MainDocumentPart;
            mainPart.DocumentSettingsPart.AddExternalRelationship("http://schemas.openxmlformats.org/officeDocument/2006/relationships/attachedTemplate",
               new Uri(targetFile, UriKind.Absolute));

            mainPart.Document.Save();
        }
        File.WriteAllBytes(targetFile, documentStream.ToArray());
    }
    public static void CopyStream(Stream source, Stream target)
    {
        if (source != null)
        {
            MemoryStream mstream = source as MemoryStream;
            if (mstream != null) mstream.WriteTo(target);
            else
            {
                byte[] buffer = new byte[2048];
                int length = buffer.Length, size;
                while ((size = source.Read(buffer, 0, length)) != 0)
                    target.Write(buffer, 0, size);
            }
        }
    }
    public static void Mailmerge(string templatePath, string DestinatePath, DataRow dr, DataColumnCollection columns)
    {
        try
        {
            dotx2docx(templatePath, DestinatePath);
        }
        catch //incase the server does not support MS Office Word 2003 / 2007 / 2010
        {
            File.Copy(templatePath, DestinatePath, true);
        }
        using (WordprocessingDocument doc = WordprocessingDocument.Open(DestinatePath, true))
        {
            var allParas = doc.MainDocumentPart.Document.Descendants<DocumentFormat.OpenXml.Wordprocessing.Text>();
            Text PreItem = null;
            string PreItemConstant = null;
            bool FindSingleAnglebrackets = false;
            bool breakFlag = false;
            List<Text> breakedFiled = new List<Text>();
            foreach (Text item in allParas)
            {
                foreach (DataColumn cl in columns)
                {
                    //<Today>
                    if (item.Text.Contains("«" + cl.ColumnName + "»") || item.Text.Contains("<" + cl.ColumnName + ">"))
                    {
                        item.Text = item.Text.Replace("<" + cl.ColumnName + ">", dr[cl.ColumnName].ToString())
                                             .Replace("«" + cl.ColumnName + "»", dr[cl.ColumnName].ToString());
                        FindSingleAnglebrackets = false;
                        breakFlag = false;
                        breakedFiled.Clear();
                    }
                    else if //<Today
                    (item.Text != null
                        && (
                                (item.Text.Contains("<") && !item.Text.Contains(">"))
                                || (item.Text.Contains("«") && !item.Text.Contains("»"))
                            )
                        && (item.Text.Contains(cl.ColumnName))
                    )
                    {
                        FindSingleAnglebrackets = true;
                        item.Text = global::System.Text.RegularExpressions.Regex.Replace(item.Text, @"\<" + cl.ColumnName + @"(?!\w)", dr[cl.ColumnName].ToString());
                        item.Text = global::System.Text.RegularExpressions.Regex.Replace(item.Text, @"\«" + cl.ColumnName + @"(?!\w)", dr[cl.ColumnName].ToString());
                    }
                    else if //Today> or Today
                    (
                        PreItemConstant != null
                        && (
                                (PreItemConstant.Contains("<") && !PreItemConstant.Contains(">"))
                                || (PreItemConstant.Contains("«") && !PreItemConstant.Contains("»"))
                            )
                        && (item.Text.Contains(cl.ColumnName))
                    )
                    {
                        if (item.Text.Contains(">") || item.Text.Contains("»"))
                        {
                            FindSingleAnglebrackets = false;
                            breakFlag = false;
                            breakedFiled.Clear();
                        }
                        else
                        {
                            FindSingleAnglebrackets = true;
                        }
                        if (PreItemConstant == "<" || PreItemConstant == "«")
                        {
                            PreItem.Text = "";
                        }
                        else
                        {
                            PreItem.Text = global::System.Text.RegularExpressions.Regex.Replace(PreItemConstant, @"\<" + cl.ColumnName + @"(?!\w)", dr[cl.ColumnName].ToString());
                            PreItem.Text = global::System.Text.RegularExpressions.Regex.Replace(PreItemConstant, @"\«" + cl.ColumnName + @"(?!\w)", dr[cl.ColumnName].ToString());
                        }
                        if (PreItemConstant.Contains("<") || PreItemConstant.Contains("«")) // pre item is like '[blank]«'
                        {
                            PreItem.Text = PreItem.Text.Replace("<", "");
                            PreItem.Text = PreItem.Text.Replace("«", "");
                        }
                        if (item.Text.Contains(cl.ColumnName + ">") || item.Text.Contains(cl.ColumnName + "»"))
                        {
                            item.Text = global::System.Text.RegularExpressions.Regex.Replace(item.Text, @"(?<!\w)" + cl.ColumnName + @"\>", dr[cl.ColumnName].ToString());
                            item.Text = global::System.Text.RegularExpressions.Regex.Replace(item.Text, @"(?<!\w)" + cl.ColumnName + @"\»", dr[cl.ColumnName].ToString());

                        }
                        else
                        {
                            item.Text = global::System.Text.RegularExpressions.Regex.Replace(item.Text, @"(?<!\w)" + cl.ColumnName + @"(?!\w)", dr[cl.ColumnName].ToString());
                        }
                    }
                    else if (FindSingleAnglebrackets && (item.Text.Contains("»") || item.Text.Contains(">")))
                    {
                        item.Text = global::System.Text.RegularExpressions.Regex.Replace(item.Text, @"(?<!\w)" + cl.ColumnName + @"\>", dr[cl.ColumnName].ToString());
                        item.Text = global::System.Text.RegularExpressions.Regex.Replace(item.Text, @"(?<!\w)" + cl.ColumnName + @"\»", dr[cl.ColumnName].ToString());
                        item.Text = global::System.Text.RegularExpressions.Regex.Replace(item.Text, @"^\s*\>", "");
                        item.Text = global::System.Text.RegularExpressions.Regex.Replace(item.Text, @"^\s*\»", "");
                        FindSingleAnglebrackets = false;
                        breakFlag = false;
                        breakedFiled.Clear();
                    }
                    else if (item.Text.Contains("<") || item.Text.Contains("«")) // no ColumnName
                    {

                    }
                } //end of each columns
                PreItem = item;
                PreItemConstant = item.Text;
                if (breakFlag
                    || (item.Text.Contains("<") && !item.Text.Contains(">"))
                    || (item.Text.Contains("«") && !item.Text.Contains("»"))
                   )
                {
                    breakFlag = true;
                    breakedFiled.Add(item);
                    string combinedfiled = "";
                    foreach (Text t in breakedFiled)
                    {
                        combinedfiled += t.Text;
                    }
                    foreach (DataColumn cl in columns)
                    {
                        //<Today>
                        if (combinedfiled.Contains("«" + cl.ColumnName + "»") || combinedfiled.Contains("<" + cl.ColumnName + ">"))
                        {
                            //for the first part, remove the last '<' and tailing content
                            breakedFiled[0].Text = global::System.Text.RegularExpressions.Regex.Replace(breakedFiled[0].Text, @"<\w*$", "");
                            breakedFiled[0].Text = global::System.Text.RegularExpressions.Regex.Replace(breakedFiled[0].Text, @"<\w*$", "");

                            //remove middle parts
                            foreach (Text t in breakedFiled)
                            {
                                if (!t.Text.Contains("<") && !t.Text.Contains("«") && !t.Text.Contains(">") && !t.Text.Contains("»"))
                                {
                                    t.Text = "";
                                }
                            }

                            //for the last part(as current item), remove leading content till the first '>' 
                            item.Text = global::System.Text.RegularExpressions.Regex.Replace(item.Text, @"^\s*\>", dr[cl.ColumnName].ToString());
                            item.Text = global::System.Text.RegularExpressions.Regex.Replace(item.Text, @"^\s*\»", dr[cl.ColumnName].ToString());

                            FindSingleAnglebrackets = false;
                            breakFlag = false;
                            breakedFiled.Clear();
                            break;
                        }
                    }
                }
            }//end of each item
            #region go through footer
            MainDocumentPart mainPart = doc.MainDocumentPart;
            foreach (FooterPart footerPart in mainPart.FooterParts)
            {
                Footer footer = footerPart.Footer;
                var allFooterParas = footer.Descendants<Text>();
                foreach (Text item in allFooterParas)
                {
                    foreach (DataColumn cl in columns)
                    {
                        if (item.Text.Contains("«" + cl.ColumnName + "»") || item.Text.Contains("<" + cl.ColumnName + ">"))
                        {
                            item.Text = (string.IsNullOrEmpty(dr[cl.ColumnName].ToString()) ? " " : dr[cl.ColumnName].ToString());
                            FindSingleAnglebrackets = false;
                        }
                        else if (PreItem != null && (PreItem.Text == "<" || PreItem.Text == "«") && (item.Text.Trim() == cl.ColumnName))
                        {
                            FindSingleAnglebrackets = true;
                            PreItem.Text = "";
                            item.Text = (string.IsNullOrEmpty(dr[cl.ColumnName].ToString()) ? " " : dr[cl.ColumnName].ToString());
                        }
                        else if (FindSingleAnglebrackets && (item.Text == "»" || item.Text == ">"))
                        {
                            item.Text = "";
                            FindSingleAnglebrackets = false;
                        }
                    }
                    PreItem = item;
                }
            }
            #endregion

            #region replace \v to new Break()
            var body = doc.MainDocumentPart.Document.Body;

            var paras = body.Elements<Paragraph>();
            foreach (var para in paras)
            {
                foreach (var run in para.Elements<Run>())
                {
                    foreach (var text in run.Elements<Text>())
                    {
                        if (text.Text.Contains("MS_Doc_New_Line"))
                        {
                            string[] ss = text.Text.Split(new string[] { "MS_Doc_New_Line" }, StringSplitOptions.None);
                            text.Text = text.Text = "";
                            int n = 0;
                            foreach (string s in ss)
                            {
                                n++;
                                run.AppendChild(new Text(s));
                                if (n != ss.Length)
                                {
                                    run.AppendChild(new Break());
                                }
                            }
                        }
                    }
                }
            }
            #endregion

            doc.MainDocumentPart.Document.Save();
        }
    }
    public static void MergeDocuments(params string[] filepaths)
    {

        //filepaths = new[] { "D:\\one.docx", "D:\\two.docx", "D:\\three.docx", "D:\\four.docx", "D:\\five.docx" };
        if (filepaths != null && filepaths.Length > 1)

            using (WordprocessingDocument myDoc = WordprocessingDocument.Open(@filepaths[0], true))
            {
                MainDocumentPart mainPart = myDoc.MainDocumentPart;

                for (int i = 1; i < filepaths.Length; i++)
                {
                    string altChunkId = "AltChunkId" + i;
                    AlternativeFormatImportPart chunk = mainPart.AddAlternativeFormatImportPart(
                        AlternativeFormatImportPartType.WordprocessingML, altChunkId);
                    using (FileStream fileStream = File.Open(@filepaths[i], FileMode.Open))
                    {
                        chunk.FeedData(fileStream);
                    }
                    DocumentFormat.OpenXml.Wordprocessing.AltChunk altChunk = new DocumentFormat.OpenXml.Wordprocessing.AltChunk();
                    altChunk.Id = altChunkId;
                    //new page, if you like it...
                    mainPart.Document.Body.AppendChild(new Paragraph(new Run(new Break() { Type = BreakValues.Page })));
                    //next document
                    mainPart.Document.Body.InsertAfter(altChunk, mainPart.Document.Body.Elements<Paragraph>().Last());
                }
                mainPart.Document.Save();
                myDoc.Close();
            }
    }

And use them like this:

DataTable dt = new DataTable();
dt.Columns.Add("Date");
dt.Columns.Add("Today");
dt.Columns.Add("Addr1");
dt.Columns.Add("Addr2");
dt.Columns.Add("PreferContact");
dt.Columns.Add("TenantName");
//......

DataRow nr = dt.NewRow();
nr["Date"] = DateTime.Now.ToString("dd/MM/yyyy");
                nr["Today"] = DateTime.Now.ToString("dd/MM/yyyy");
//......
dt.Rows.Add(nr);


string sourceFile = "c:\my_template.docx"; //this is where you store your template
string filePath = "c:\final.docx"; //this is where your result file locate
Mailmerge(sourceFile, filePath, nr, dt.Columns);

Your template(c:\my_template.docx) will be just like normal .docx file, and you need to specify your fields in it:

<field>

So, your template(c:\my_template.docx) should be like:

<Today> 

<DebtorName>
<DebtorADDR>
<DebtorEmail>


Dear <Dear>,

Congratulations on yourr property <PlanNo>  <BuildAddress>. Your unit number is <LotNo> ...............

In addition, if some of your fields contain line breaks, use this:

nr["Address"] = my_address_text_contains_line_breaks.Replace(Environment.NewLine, "MS_Doc_New_Line");