A visual canvas for matching, built for accuracy
Connect steps on a canvas, or let the AI assistant draft the whole process from your column names. Every step can be tuned, tried on a sample, and run when you are ready.
Data preparation
Clean and reshape both files before anything is compared: normalize text, set column types, drop rows that should never match. The AI assistant suggests transformations from your column names and sample data.
- Custom regex transforms
- Type definitions for numbers and dates
- Row Filter: keep or remove rows by value
AI extraction and enrichment
Pull structured attributes out of unstructured text or files, or generate fields that are not in the raw data. Extract product names from descriptions, parse addresses, classify records, or read fields from PDFs and images.
- Structured output from free text
- Images, PDFs and URLs as input
- Optional search, maps and URL tools for enrichment
Embeddings and similarity
Turn text into vectors for semantic matching that goes beyond keyword overlap, then score candidate pairs. Similar Items finds the close pairs directly, so a large file is never turned into every possible combination first.
- Choose the embedding size, from fast to maximum
- Embed several columns together with a template
- Per-field thresholds
Rule-based matching
Exact keys, string rules and numeric tolerance for the cases where meaning is not the question: codes, identifiers, prices, dates. Cheap rules run first, so the AI only sees pairs that survive them.
- Join on exact keys
- Fuzzy Score: token set, token sort, Levenshtein, Jaro-Winkler
- Numeric filter with percentage tolerance
- Score Filter on any score column
AI checks
For borderline pairs, an AI check reads both records against your own prompt and returns an answer you define. Answers can send a pair on as a match, hold it for a person, or reject it.
- Custom prompts and answer choices
- Optional reason recorded per pair
- Standard or boosted intelligence
- Queued processing: a lower price for a slower response
Human review
Put people in the loop where it matters. Uncertain pairs become a queue of questions with a deadline. Reviewers answer only the items assigned to them, and everyone can see afterwards who answered what.
- Reviewer role with access to assigned items only
- Assign by list or by percentage, with default reviewers
- E-mail when a review needs attention
One-to-one matching and grouping
Find items matched more than once and keep the best partner, or turn accepted pairs into groups of identical items across both files, one row per item.
- Resolve Matches: flag or keep the best partner
- Group Items from accepted pairs
- Remove duplicates inside one file
Estimate, sample, stop
See the most a run can use before it starts, try a sample on your own data, and stop a run at any time. Credits held for work that did not happen come back to your balance.
- Estimate before every run
- Sample runs on your real rows
- Stop a run and keep only what was used
Team, API and agents
Invite people with roles that fit their work, and run the same canvases from code. The API is described in an open OpenAPI document, and an installable skill teaches AI agents to use it with the same credit safeguards.
- Roles: owner, admin, member, viewer, reviewer
- API tokens with usage limits
- OpenAPI description and agent skill
Export
Download results as a clean CSV with the columns you choose, in the order you choose, from either file and from the match itself.
- Custom column selection and order
- Scores and review answers included
- One file per Export step
Start matching your data today
Start with 100 free credits. No credit card required.
Get started free →