Deduplicate your CSV in one click. Try it: https://dedupe.it
Watch a demo:
Dirty datasets suck. Common problems:
- Same person signs up with business and personal email
- Same company gets added to your CRM twice with slightly different naming, ex. "Amazon" vs. Amazon.com"
- Two overlapping datasets get merged - ex. website visitors and demo form submissions
I've wasted tons of time cleaning datasets with Excel formulas and SQL, but hand-crafting deterministic rules is tedious.
After trying "AI-powered" solutions like the AWS Entity Resolution, I was frustrated to find that I still had to map fields and hand-tune "confidence scores".
Dedupe It obviates manual configuration by using LLM vector embeddings & chat completions.
The algorithm:
- Perform an LLM vector embedding of each record
- For each record, perform vector similarity search to identify its nearest neighbors (typically 2-5 neighbors)
- For each nearest neighbor, ask an LLM to determine whether it's a duplicate of the record under consideration
- If we identify two records as duplicates, we group them together
- For each duplicate group, we have our LLM propose a merged record
Here are some advantages of this approach:
- Unlike a pure NxN record comparison, scales approximately linearly in time and cost
- Unlike a pure vector clustering approach, avoids the hard problem of picking a cluster cutoff
- Utilizes LLM chat completions to make fine-grained comparisons. Also makes it possible to add custom guidelines, ex. "consider international subsidiaries to be distinct entities".
- Incremental / stream deduplication: Persist the vector database and groupings to deduplicate new records without reprocessing the whole dataset.
- Custom rules: Allow specifying custom guidelines, either fuzzy (ex. "consider international subsidiaries to be distinct entities") or deterministic (always use the values from the last updated record when merging).
- Integrations: Excel, Google Sheets, Google Forms, Salesforce, Hubspot, Snowflake, etc.
NOTE: Some of these features are already in the works. If you have requests or want early access, email founders@snowpilot.com or fill out the form on the demo site (https://dedupe.it)
Frontend: see instructions in frontend/README.md
Backend: see instructions in backend/README.md
Contributions are welcome. We don't currently have a formal process for contribs; feel free to simply open an issue or PR.
Dedupe It is open-source software, available under the Apache 2.0 license.
If you intend to use Dedupe It for commercial purposes, we simply ask that you include a copy of the original license in your code, and email the authors (founders@snowpilot.com) with a short explanation of how you intend to use Dedupe It.
What does Dedupe It cost to run?
With no modifications, this implementation will cost you ~$8.40 per 1k records. This will depend on the size of your records, so please benchmark your costs on a small, representative sample of data before attempting to run on a large dataset.
NOTE: We're working on optimizations to reduce the cost 100x. If you need to run deduplication on a larger dataset (100k+ rows), please email us: founders@snowpilot.com
When should I use Dedupe It?
Dedupe It is built for fuzzy deduplication of structured data. In other words, it's a good fit if your data:
- Fits into a spreadsheet / CSV
- Might have duplicates
- The duplicates are not exact (ex. "Michael B. Schmidt, michael.schmidt@acme.com" and "mike schmidt, mike@acme.com")
This is typically the case if:
- Your data comes from manual entry, like someone filling out a form
- Your data is a combination of multiple sources, each of which has different fields or formatting
Do you have performance benchmarks?
We'll release some more detailed head-to-head and benchmarks in the future.
Anecdotally, on the datasets that we've tested (people, companies, and products):
- Rate of false negatives (not catching a pair of duplicates): <5%
- Rate of false positives (mistakenly identifying distinct records as duplicates): vanishingly small
Why should I use Dedupe It over alternative X?
As of today, our main value proposition is simplicity.
There is no other fuzzy deduplication solution that works out-of-the-box without manual configuration.
That said, this implementation is a minimum viable product. Shortcomings of Dedupe It today include:
- Relatively expensive (~$8.40/1k records). NOTE: we will be bringing this down 100x over the next month or two.
- No incrementality. NOTE: we have incremental versions working locally.
- No integrations. NOTE: on the roadmap.
- No custom rules. NOTE: on the roadmap.
If you need any of those features, please contact us.
In the meantime, here are some popular alternatives:
- Enterprise-ready: AWS Entity Resolution
- Open-source: Zingg
- Integrated with Salesforce: DataGroomr
Can you add feature X?
Probably! Add a thread in the discussions section on Github, email us at founders@snowpilot.com.
Dedupe It is built by the Snowpilot team
To contact us, email founders@snowpilot.com
Cheers!