RoloDexter

The universal contact field mapper.

Route messy, inconsistent contact data from any source to a clean, canonical schema.

---

RoloDexter is an open-source Python and TypeScript library that maps messy, inconsistent contact field names, like HubSpot's firstname or a CSV's Column A, onto one canonical schema of 62 fields. It resolves each header through a four-layer pipeline, exact, normalized, fuzzy, heuristic, scores every match's confidence, and normalizes the matched values so the output is ready to store.

Packages In This Repository

RoloDexter is maintained as a dual-package repository:

Ecosystem	Package	Source	Package metadata	Publish target
Python	rolodexter	src/rolodexter	pyproject.toml	PyPI
JavaScript / TypeScript	rolodexter	packages/js/src	packages/js/package.json	npm

The Python package remains the canonical implementation and owns the shared patterns.json alias table. The NPM package lives under packages/js, syncs that alias table during build, and has its own README, TypeScript sources, tests, package metadata, and publish workflow.

The Problem

Every CRM, email platform, and CSV export uses different field names for the same data:

Service	First Name	Phone	Company
HubSpot	firstname	mobilephone	company
Salesforce	FirstName	MobilePhone	Company
Mailchimp	FNAME	PHONE	COMPANY
Google CSV	Given Name	Phone 1 - Value	Organization 1 - Name
Random CSV	Column A	Column B	Column C

The Solution

from rolodexter import ContactMapper

mapper = ContactMapper()

result = mapper.map_payload({
    "fname": "jane",
    "surname": "doe",
    "mobile": "+1-650-253-0000",
    "employer": "Tech Corp",
    "Column 1": "[email protected]",  # auto-detected by value shape
})

print(result.normalized)
# {
#     "first_name": "Jane",
#     "last_name": "Doe",
#     "phone": "+16502530000",
#     "company": "Tech Corp",
#     "email": "[email protected]"
# }

Installation

Python

# Core (phonenumbers + nameparser)
pip install rolodexter

# With fuzzy matching for typo recovery
pip install rolodexter[fuzzy]

# With on-demand i18n cache generation dependencies (40 languages)
pip install rolodexter[i18n-generate]

# Everything
pip install rolodexter[all]

# Development
pip install rolodexter[dev]

JavaScript / TypeScript

npm install rolodexter

The NPM package source lives in packages/js. For local development:

cd packages/js
npm install
npm test

Features

🔎 Pre-Flight Profiling: "What Will I Lose?"

# See what a file maps to before you write anything
rolodexter profile contacts.csv

# Same report as JSON, for scripting
rolodexter profile contacts.csv --json

# Skip value-level normalization for a faster pass on a big file
rolodexter profile huge.csv --no-normalize --max-rows 5000

profile never writes mapped output. It reports the match rate, which canonical fields got populated, which headers went unmapped, and categorized warning counts, so you can see what an import will drop or flag before running map for real. It accepts the same --region, --languages, --min-confidence, and --override flags as map.

🎯 Four-Layer Matching Pipeline

Every field runs through the strategy chain in priority order:

1. Exact Match: O(1) lookup against 600+ known aliases across 62 canonical fields
2. Normalized Match: handles CamelCase, dot.path, space → underscore, vendor and address prefixes (hs_email, Billing City), a form question's leading "Your" (Your company name), and similar variations
3. Fuzzy Match: rapidfuzz catches typos like "phne_nmbr" → phone
4. Heuristic Match: regex detects emails, phones, URLs, postal codes by data shape

📊 Confidence Scoring

Every match comes with a confidence score (0.0–1.0):

match = mapper.identify("fname")
# FieldMatch(original='fname', canonical='first_name', confidence=1.0, strategy='exact')

match = mapper.identify("phne")
# FieldMatch(original='phne', canonical='phone', confidence=0.7, strategy='fuzzy')

match = mapper.identify("Column X", value="[email protected]")
# FieldMatch(original='Column X', canonical='email', confidence=0.6, strategy='heuristic')

Per-Caller Field Overrides

For vendor-specific or account-level field names that won't be in the standard alias table:

mapper = ContactMapper(
    overrides={
        "MMERGE6": "company",   # Mailchimp custom merge field
        "cf_lead_score": "tags",
    }
)

📱 Phone Extraction

# Extract phones embedded in arbitrary string values
result = mapper.map_payload(
    {"notes": "call me at +1-650-253-0000 or +44 20 7946 0958"},
    extract_embedded_phones=True,
)
print(result.get_all_phones())
# ['+16502530000', '+442079460958']

🗂️ Tags / List Fields

Fields like tags are automatically list-normalised: comma-separated strings, JSON arrays, and Python lists all collapse to a clean list:

result = mapper.map_payload({"tags": "vip, newsletter, beta"})
print(result.normalized["tags"])
# ['vip', 'newsletter', 'beta']

🌍 On-Demand i18n (40 Languages)

English ships by default. Install rolodexter[i18n-generate] to generate any of 40 supported language caches with the i18n CLI or API, then pass those languages to ContactMapper; runtime loading is cache-only and never translates during mapper construction:

from rolodexter import ContactMapper

# Load Spanish aliases from a generated cache
mapper = ContactMapper(languages=["es"])
result = mapper.map_payload({"correo_electronico": "[email protected]"})
print(result.normalized["email"])  # [email protected]

# CLI: generate and cache all 40 languages
python -m rolodexter.i18n

# Or specific languages
python -m rolodexter.i18n --languages es,fr,de

# Bound network behavior during generation
python -m rolodexter.i18n --languages es,fr --timeout 10 --retries 1 --workers 4

# List supported languages
python -m rolodexter.i18n --list

Supported: Spanish, French, German, Portuguese, Italian, Dutch, Polish, Romanian, Turkish, Russian, Japanese, Chinese (Simplified), Korean, Arabic, Hindi, Swedish, Danish, Norwegian, Finnish, Czech, Ukrainian, Greek, Hungarian, Thai, Vietnamese, Indonesian, Malay, Hebrew, Bulgarian, Croatian, Slovak, Slovenian, Serbian, Lithuanian, Latvian, Estonian, Catalan, Filipino, Swahili, Afrikaans.

🧹 Value Normalization

Automatic cleanup on matched fields:

• Phone → E.164 format via libphonenumber (+16502530000)
• Email → lowercase, trimmed
• Names → title case with particle awareness ("jane van der berg" → "Jane van der Berg")
• Addresses → excess whitespace collapsed, title-cased
• Tags → normalized to list[str]
• Dates → ISO-8601 for birthday, created_at, updated_at, last_contacted. Ambiguous values (03/04/2024, a two-digit year) are left unchanged and reported as a warning instead of guessed.
• Country → ISO 3166-1 alpha-2 ("Deutschland" → "DE")
• State / Province → 2-letter US state or Canadian province code ("california" → "CA")

📦 Batch & Streaming

results = mapper.map_batch([contact1, contact2, contact3, ...])

# Constant-memory streaming for huge CSV/JSONL exports:
import csv
with open("contacts.csv") as fh:
    for result in mapper.map_stream(csv.DictReader(fh)):
        save(result.normalized)

# Preview import readiness without retaining mapped rows:
profile = mapper.profile(contacts, max_rows=1_000)
print(profile.explain())

🐼 DataFrames

import pandas as pd
from rolodexter import ContactMapper

df = pd.read_csv("hubspot_export.csv")
clean = ContactMapper().map_dataframe(df)   # pip install rolodexter[pandas]
# Columns renamed to canonical fields, values normalized, unmatched columns kept.

🖥️ Command Line

# Map a CSV/JSON/JSONL export to the canonical schema
rolodexter map contacts.csv -o clean.csv --region US

# Stream JSON Lines, drop low-confidence guesses, fail loudly
rolodexter map export.jsonl --min-confidence 0.8 --strict -o out.jsonl

# JSON/CSV output paths are bounded; JSONL output remains streaming
rolodexter map huge.jsonl --format jsonl --max-materialized-rows 100000

# Keep processing after bad rows, preserving failures in a JSONL quarantine file
rolodexter map export.jsonl --strict --on-error quarantine -o clean.jsonl

# Carry unmapped columns through instead of dropping them, and drop rows
# that share an identity key (email/phone/source id) with an earlier row
rolodexter map contacts.csv --keep-unmapped --dedupe -o clean.csv

# Force a vendor-specific column to a canonical field
rolodexter map contacts.csv --override "MMERGE3=full_address" -o clean.csv

# Save the resolved header plan, then replay it on a later import so columns
# route identically even after a patterns.json change (a mapping lockfile)
rolodexter map jan.csv --schema-out plan.json -o jan-clean.csv
rolodexter map feb.csv --schema-in plan.json  -o feb-clean.csv

# See exactly how a header resolves
rolodexter explain "Job Titel" --value CEO
# 'Job Titel' -> job_title [fuzzy, conf=0.70]

rolodexter --version
rolodexter fields        # list every canonical field

🛡️ Strict Mode, Warnings & Confidence

# Non-fatal issues are reported, never silent:
result = mapper.map_payload({"mobile": "not a phone"})
print(result.warnings)
# ("'mobile': phone value 'not a phone' could not be normalized to E.164 ...",)

# Each warning is a str subclass carrying its category, for grouping by code
# instead of matching substrings of the message text:
print(result.warnings[0].category)   # 'phone_normalization'

# Demand high-confidence mappings; fail loudly on any problem:
mapper = ContactMapper(strict=True, confidence_threshold=0.8)

print(result.explain())   # human-readable resolution + warnings

🗺️ Compile a Schema Once

schema = mapper.compile_schema(["First Name", "Mobile Phone", "Org"])
schema.column_map()         # {'First Name': 'first_name', 'Mobile Phone': 'phone', 'Org': 'company'}
schema.apply(row)           # reuse the resolved plan per row

# Save the plan and replay it later so columns route identically, even after
# a patterns.json update (the "map --schema-out" / "--schema-in" CLI flags
# do this for you):
plan = schema.to_dict()
schema2 = MappingSchema.from_dict(plan, mapper)

📈 Rich Diagnostics

result = mapper.map_payload(data)

print(result.match_rate)        # 0.857
print(result.matched_count)     # 6
print(result.unmatched_count)   # 1
print(result.get_all_phones())  # ['+16502530000']
print(result.get_all_emails())  # ['[email protected]']
print(result.get_identity_keys())
# ['email:[email protected]', 'phone:+16502530000', 'source:hubspot:123']
print(result.to_dict())         # Full JSON-serializable report

🔢 Nested Payload Support

# Flatten one level of nesting with depth=2
result = mapper.map_payload(
    {"contact": {"fname": "Jane", "lname": "Doe"}},
    depth=2,
)
# Accesses "contact.fname" and "contact.lname"

API Reference

ContactMapper

ContactMapper(
    *,
    patterns=None,             # Custom pattern dict (overrides built-in)
    patterns_path=None,        # Path to a custom patterns.json file
    normalize=True,            # Apply value normalization after mapping
    strategies=None,           # Override the default strategy pipeline
    languages=None,            # None=English only | "es" | ["es","fr"] | "all"
    overrides=None,            # Extra alias→canonical mappings {"MMERGE6": "company"}
    default_region="US",       # ISO-3166 region for phone parsing/E.164
    strict=False,              # Raise NormalizationError on any warning
    confidence_threshold=0.0,  # Drop matches below this confidence to unmapped
    header_cache_max_size=4096,# Bound header-resolution cache; None=unbounded
)

Methods:

Method	Description
identify(header, *, value)	Resolve a single header to a FieldMatch
trace_header(header, *, value)	Every pipeline layer's verdict, in order
map_payload(payload, *, depth, ...)	Normalize an entire dict → MappingResult
map_batch(payloads, *, ...)	Process a list of payloads → list[MappingResult]
map_stream(iterable, *, ...)	Lazily yield results (constant memory)
profile(iterable, *, max_rows, ...)	Aggregate import-readiness diagnostics
compile_schema(headers)	Resolve headers once → reusable MappingSchema
map_dataframe(df)	Rename/normalize a pandas DataFrame
clear_cache()	Clear cached header-resolution verdicts
cache_info()	Inspect header cache size/configuration
registry	Access the underlying PatternRegistry

map_dataframe() guarantees unique output labels. It skips reserved <canonical>__N names when an unmatched source column already uses one, and rejects duplicate input labels because they cannot be renamed without ambiguity.

FieldMatch

FieldMatch(
    original='fname',
    canonical='first_name',
    confidence=1.0,
    strategy='exact',      # 'exact' | 'normalized' | 'fuzzy' | 'heuristic' | 'none'
    is_matched=True,
)

MappingResult

Attribute / Method	Type	Description
normalized	dict	Canonical key → cleaned value
unmapped	dict	Fields that couldn't be resolved
field_matches	tuple[FieldMatch, ...]	Full match detail for every input field
match_rate	float	Fraction of fields successfully matched
matched_count	int	Count of matched fields
unmatched_count	int	Count of unmatched fields
warnings	tuple[str, ...]	Non-fatal issues (failed E.164, dropped matches)
get_match(header)	FieldMatch | None	O(1) lookup of the match for an input header
get_all_phones()	list[str]	All phone values across all phone-adjacent fields
get_all_emails()	list[str]	All email values, flattened and deduplicated
get_identity_keys()	list[str]	Stable email/phone/source keys for deduplication
explain()	str	Human-readable resolution + warnings summary
to_dict()	dict	Full JSON-serializable report

MappingProfile

ContactMapper.profile() consumes any iterable in constant memory and reports rows and fields seen, aggregate match rate, canonical-field counts, unmapped header counts, strategy usage, and categorized warnings. Use max_rows for a bounded preview; the next iterator item is left untouched.

profile = mapper.profile(rows, max_rows=1_000, confidence_threshold=0.8)
profile.match_rate
profile.warning_counts
profile.to_dict()   # JSON-ready
profile.explain()   # compact report for logs or a CLI

CanonicalField

Enum of all 62 canonical fields. Inherits from str for JSON compatibility:

from rolodexter import CanonicalField

assert CanonicalField.EMAIL == "email"
assert CanonicalField.PHONE.value == "phone"

first_name · last_name · full_name · middle_name · nickname · prefix · suffix · email · phone · home_phone · work_phone · fax · whatsapp · website · company · job_title · department · industry · address_line1 · address_line2 · city · state · postal_code · country · full_address · linkedin · twitter · facebook · instagram · github · youtube · tiktok · discord · telegram · lead_status · lifecycle_stage · email_opt_out · tags · source · utm_parameters · score · owner · birthday · age · created_at · updated_at · last_contacted · revenue · currency · message · subject · company_size · notes · metadata · gender · timezone · language_preference · referrer_url · source_id · source_service · subscribed · verified · unknown

Custom Patterns

custom = {
    "fields": {
        "first_name": ["fname", "given", "nombre"],
        "loyalty_tier": ["tier", "vip_level", "membership"],
    }
}

mapper = ContactMapper(patterns=custom)

Custom pattern data is validated when the registry is constructed. fields must be an object whose values are lists of non-empty aliases; malformed files raise PatternLoadError with the failing section instead of partially building an index.

Repository Layout

rolodexter/
├── pyproject.toml              # Python/PyPI package metadata
├── src/rolodexter/             # Python package source
├── tests/                      # Python tests
├── packages/js/package.json    # JavaScript/NPM package metadata
├── packages/js/src/            # TypeScript package source
├── packages/js/test/           # JavaScript package tests
└── scripts/                    # Cross-language release/parity probes and sweep

Python Package Architecture

src/rolodexter/
├── __init__.py      # Public API
├── __main__.py      # CLI: rolodexter map / explain / fields
├── core.py          # ContactMapper, PatternRegistry, strategies, normalizers
├── _phone.py        # E.164 phone parser (wraps libphonenumber)
├── i18n.py          # On-demand i18n generator (40 languages, cached)
└── patterns.json    # Master alias table (600+ aliases, 62 canonical fields)

Contributing

git clone https://github.com/LunarWerxs/RoloDexter.git
cd rolodexter
pip install -e ".[dev]"
pytest

The public API of both packages is pinned by golden snapshots in scripts/api_golden/: one text file per export, plus a hash per export and member that also covers every type it depends on. npm run check:parity (run by CI) fails when either surface changes and labels each difference: a removed name is BREAKING, a changed one POTENTIALLY_BREAKING, a new one POTENTIALLY_NON_BREAKING, with a suggested version bump. After a deliberate API change, review the report and accept it:

python scripts/api_snapshot.py                   # report; exits 1 on any change
python scripts/api_snapshot.py --update-goldens  # rewrite only the affected goldens

The Python goldens are rendered with the Python version CI pins (3.14): inspect.signature and dataclass-generated __init__ text can differ between versions, so run the snapshot (and npm pack, whose prepack runs it) on 3.14, or pass --lang js to check only the Node half.

FAQ

Is RoloDexter free? Yes. RoloDexter is MIT-licensed and free to use in commercial and open-source projects, both the Python package on PyPI and the JavaScript/TypeScript package on npm. There is no paid tier, account, or API key involved: it installs as a normal dependency and runs entirely inside your own process.

Does it work offline? Yes, for field mapping and value normalization, both run entirely locally with no network calls. The one exception is the optional i18n alias generator (rolodexter[i18n-generate]), which calls Google Translate to build language caches; that step is explicit, one-time, and the caches it produces are then loaded offline with no further network access.

What are the system requirements? The Python package needs Python 3.10 or newer (tested through 3.14) plus its two core dependencies, phonenumbers and nameparser; optional extras add rapidfuzz for fuzzy matching, pandas for DataFrame support, or translation libraries for i18n generation. The JavaScript/TypeScript package needs Node.js 20 or newer. Neither requires a database or server.

How is it different from Zapier's or Make's field mapping? Zapier and Make connect two apps by having you manually map each field between them for every workflow you build, by hand, per pill. RoloDexter instead resolves headers automatically against a built-in table of 600+ aliases across 62 canonical fields, so a HubSpot export and a random CSV both land on the same schema without you wiring each column yourself.

How is it different from an ETL platform like Talend or Alteryx? Talend and Alteryx are full ETL platforms with schema mapping as one feature among many; Talend Data Fabric alone is priced for enterprise budgets, commonly six figures a year. RoloDexter is a free, MIT-licensed library that does one job, contact field mapping and value normalization, embedded directly in your own Python or Node code with no platform to manage.

How is it different from using phonenumbers or nameparser directly? phonenumbers and nameparser, two libraries RoloDexter depends on, each solve one field type: phone parsing and name parsing. Neither tells you which of fifty CSV columns is a phone number, or that mobilephone and mobile both mean phone. RoloDexter runs header identification across all 62 canonical fields first, then hands the matched values to libraries like these for normalization.

What data formats does it support? CSV, JSON, and JSON Lines files through the CLI (rolodexter map), plus plain Python dicts, lists, and streaming iterables through the library API, and pandas DataFrames with the pandas extra installed. Output can be written as CSV, JSON, or streamed JSONL, and a --schema-out/--schema-in lockfile replays a resolved header plan on later imports.

Is my data sent anywhere? No. Field mapping and value normalization run entirely inside your own Python or Node process; RoloDexter makes no network calls and has no server component. The only exception is the optional, explicitly-invoked i18n alias generator, which sends field names, not your contact data, to Google Translate to build a local language cache.

License

MIT, see LICENSE.

---

Made by LunarWerx Studios, the team also behind RepoYeti, AgentHydra, and SageThumbs.