CSV/Excel Merger skill

CSV/Excel Merger

Combine tabular files into one clean output without silently losing, duplicating, or corrupting rows.

Workflow

Copy this checklist and track progress:

- [ ] 1. Profile inputs
- [ ] 2. Choose append vs join
- [ ] 3. Map columns and normalize keys
- [ ] 4. Merge and resolve conflicts
- [ ] 5. Verify row math
- [ ] 6. Write output and report

1. Profile inputs. Run the bundled profiler first; it reports encoding, delimiter, rows, headers, candidate keys, and header overlap without changing anything:

   python scripts/profile_inputs.py file1.csv file2.xlsx
   

   Excel files are profiled per sheet. Confirm with the user which sheets count if a workbook has more than one.

2. Choose the operation. This is the decision that most often goes wrong:

   • Append (stack) - same kind of records from different sources or periods (Jan + Feb exports, three lead lists). Use pd.concat, then dedupe.
   • Join (enrich) - different facts about the same entities (contacts + their deal values). Use pd.merge on a key.
   • Unsure? If the files share most columns, append. If they share only an ID column, join.

3. Map columns and normalize keys. Build an explicit {original: unified} rename map per file (see references/merge_strategies.md for common variants) and show it to the user when any match is fuzzy. Normalize key columns before dedupe or join: strip whitespace, lowercase emails, strip non-digits from phones, unify date formats. Without this, [email protected] and [email protected] survive as two people.

4. Merge. Read every file with dtype=str so IDs, ZIP codes, and phone numbers keep leading zeros, then convert specific columns afterward.

   import pandas as pd
   
   frames = []
   for path, rename in [("jan.csv", {"E-mail": "email"}), ("feb.xlsx", {"Email Address": "email"})]:
       df = (pd.read_excel(path, dtype=str) if path.endswith((".xlsx", ".xls"))
             else pd.read_csv(path, dtype=str, encoding="utf-8-sig",  # use the profiler's encoding
                           keep_default_na=False))
       df = df.rename(columns=rename)
       df["email"] = df["email"].str.strip().str.lower()
       df["source_file"] = path          # lineage for every row
       frames.append(df)
   
   combined = pd.concat(frames, ignore_index=True, sort=False)
   # Blank keys are not duplicates of each other: set them aside before deduping.
   has_key = combined["email"].fillna("") != ""
   # Later files win: list the most recent source last, then keep="last".
   deduped = combined[has_key].drop_duplicates(subset=["email"], keep="last")
   no_key = combined[~has_key]
   merged = pd.concat([deduped, no_key], ignore_index=True)
   

   For a join, make pandas enforce the relationship you expect so a duplicate key raises instead of multiplying rows:

   out = pd.merge(contacts, deals, on="email", how="left",
                  validate="one_to_one", indicator=True)
   unmatched = out[out["_merge"] == "left_only"]
   

   Conflict strategies (keep first/last/most complete, combine fields, flag for review) are in references/merge_strategies.md.

5. Verify before reporting. Never hand back a merge without checking it:

   rows_in = sum(len(f) for f in frames)
   assert len(merged) > 0, "merge produced an empty frame"
   assert len(merged) <= rows_in, "more rows out than in: check the join keys"
   assert deduped["email"].is_unique, "duplicate keys remain after dedupe"
   print(f"in={rows_in} out={len(merged)} removed={rows_in - len(merged)} blank_keys={len(no_key)}")
   print(merged["source_file"].value_counts())
   

   Spot-check three removed duplicates by hand against the source files; the asserts prove the math, not that the right row won.

6. Write output and report. Use the layout in references/output_template.md.

   • CSV for Excel users: to_csv(path, index=False, encoding="utf-8-sig") (the BOM makes Excel read accents correctly).
   • Excel: to_excel(path, index=False) with openpyxl installed. A sheet holds at most 1,048,576 rows; split or use CSV/Parquet beyond that.
   • Also write conflicts_review.csv or unmatched.csv when those sets are non-empty.

pandas version notes

Current pandas is 3.x (Python 3.11+). Differences that affect merges:

• Text columns default to the str dtype, not object. Check pd.api.types.is_string_dtype(col) instead of dtype == object.
• Copy-on-Write is always on. Chained assignment such as df[col][mask] = x never updates df (pandas only warns); use df.loc[mask, col] = x.
• Parsed datetimes default to microsecond resolution. Call .dt.as_unit("ns") before casting to integers if something downstream expects nanoseconds.
• pd.read_excel(..., engine="calamine") (needs python-calamine) reads large workbooks much faster than openpyxl.

The code in this skill also runs on pandas 2.2.

Failure modes to check for

• Row explosion on join - duplicate keys on both sides multiply rows. validate= catches it.
• Leading zeros lost - reading without dtype=str turns 01234 into 1234.
• Excel-mangled values - long IDs already shown as 1.23E+15 or dates already reformatted in the source file cannot be recovered by pandas; flag them.
• Header rows not on line 1 - exports with a title block need skiprows= or header=.
• Mixed encodings - one file in cp1252 among UTF-8 files shows up as é artifacts. The profiler reports the encoding per file.
• Silent column drops - a column present in only one file becomes mostly empty after append. Keep it and report its completeness; never drop data without saying so.
• Large files (over a few hundred MB) - read with chunksize= or use Polars/DuckDB, and dedupe with a key set instead of loading everything into memory.