How this works
Every claim on this page is something you can hold us to. If any of it stops being true, this page is wrong and we want to hear about it.
Retrieval first, reasoning second
The order is fixed in code, not requested in a prompt. A language model is good at eliminating candidates from a set and bad at being the set.
Typing a code skips the first step, and stops after the third unless you ask for help choosing: the search results are a database query and no model is involved. A photo runs the first step; Help me choose runs the fourth.
- A vision pass reads the characters off your photo and nothing else. It returns per-character alternatives rather than one string, and it is forbidden from naming a part. Letting a vision model guess a part number is exactly how a lookup tool becomes confidently wrong.
- The reading is normalised. Confusable characters collapse into one class, so the search finds the part whether the photo read 6H, GH or CH. The exact characters are kept alongside the normalised form.
- Our marking database is queried for every part matching the code and the package. If that returns nothing, the search widens: prefix match first, then the confusable set. The result page says which of those happened.
- Only then does the model see anything. It is given the retrieved set and the circuit context you supplied, and its job is to rank, eliminate, and write the reasoning and the confirming test.
Over-retrieving is the correct direction to fail in. A candidate that was never retrieved can never be recovered, and one that was retrieved and argued against costs you a sentence to read.
The model works from a closed set and cannot leave it
The candidate set passed to the model is the only set of part numbers it is allowed to return. Every part number in the response is checked back against that set before you see it, and anything that was not in it is dropped.
We count the violations
The one case where no part number is returned at all is described under functional identification. That output is stored in a different table from identified candidates and rendered in a different block, so an inference can never be dressed up as an identification by an interface bug.
Confidence is a word, never a percentage
There are three bands and they are always spelled out. A number like “83 percent” would imply we had calibrated this against a large body of confirmed outcomes and could tell 83 from 79. We have not and we cannot, and printing the number anyway would be a lie told in a font.
| Band | What it means |
|---|---|
Likely | The circuit context fits this part and argues against the others. Run the test anyway. |
Possible | Nothing eliminates it, and nothing much argues for it either. This is usually where the question loop earns its place. |
Long shot | Retrieved, argued against, and kept anyway. A candidate you are never shown is a candidate you can never come back to at midnight. |
There is no red and green traffic light either. You read schematics. A coloured dot tells you nothing that a word and a sentence of reasoning do not tell you better.
Where the data comes from, and what we refuse
The corpus is built from vendor datasheet marking tables and grown by confirmations from technicians. We checked the actual terms of every other source before ingesting anything. None of them says “free for non-commercial use”. That phrasing appears nowhere.
| Source | Why |
|---|---|
| Vendor datasheet marking tables used | Published by Nexperia, onsemi, Diodes, ROHM, Vishay, TI, ST, Infineon and Toshiba so that people can identify parts. The mapping inside is fact. |
| Confirmations from technicians using WhatSMD used | Submitted to us under our own terms, and the only source that ever covers house-numbered and in-country parts. |
| Distributor APIs used | Used live for package, pinout and datasheet links where a licence allows it. None of them carries a marking-code index at all: distributors leave marking codes out deliberately, as an anti-counterfeiting measure. |
| The circulating SMD code books not used | Copyright notice, no licence grant of any kind. We use one as a private accuracy check and ship none of its rows. We have asked the author for a licence. |
| markingcodes.com, embedeo, alltransistors, s-manuals not used | One reserves all rights, one forbids manual and automated extraction, one carries a machine-readable EU opt-out and blocks our crawler by name, and one disallows exactly those paths in robots.txt. |
We build in Slovenia, so the binding rule is the EU database right rather than the American argument that facts cannot be owned. That is a real constraint and it costs us coverage on day one. It is also why the confirmation loop below is the business rather than a nice touch.
The confirmation loop, and why it pays
After every identification there is one question: “Did this fix the board?” Three buttons. Yes, no, still testing.
A yes with a named part writes a verified row: marking, package, board make and model, circuit context, confirmed part number. That is a resolution nobody else has, and it makes the next technician who meets the same code get a better answer than you did.
Confirmations earn free months. We would rather say plainly what that is than dress it up: it is training data we are choosing to pay for instead of guess at. A no is worth having too. It goes into the weekly accuracy review, which is the only way we find out where the ranking is wrong.
The ranking and the reasoning are written by an AI model
Saying so is a legal requirement and it is also just the truth. Three things on a result page come from a language model: the characters read off your photo, the order of the candidates with the sentence explaining each one, and the confirming test. Your photos and the circuit context you type are sent to our model provider to produce them. Typed search sends nothing to a model: its results come straight from our tables.
What does not come from a model: the candidate list itself, which is a database query, and the part numbers, which are filtered back against that query. Every model response is validated against a strict schema before anything is stored, and a response that fails validation is retried once and then fails loudly rather than being patched into shape.
What happens to your photos, who processes them and how to delete them is in the privacy policy. The short version: they are used for the identification and the confirmation record, and for nothing else unless you turn that on yourself.
What it is bad at
Specifically, and without hedging. If one of these is your case, save your time.
- Unmarked passives. A burnt resistor with no band and no print is not an identification problem we can solve. The circuit narrows a resistor to an order of magnitude at best, and we will tell you that rather than name a value. Find the same part elsewhere on the board or find the schematic.
- Codes nobody ever published. Our corpus is built from what vendors printed in datasheets. House numbers and in-country copies were never in there. Functional identification is a fallback for that case, not a substitute: it tells you what the part must do, not what it is.
- Thin context, which is most of the bad outcomes. “It is on a TV board” moves the ranking almost nothing. Which pin goes to which rail moves it enormously. If you cannot trace the pins, expect several candidates at the same confidence and a page full of questions.
- Package guessed from one shot. SOT-23 and SOT-323 look identical in a photo without something for scale, and getting the package wrong poisons the retrieval before any reasoning happens. Put a probe tip or a 0603 in the frame.
- Second sources with a shared code. Where two manufacturers use the same marking for parts with different pinouts, the circuit sometimes separates them and sometimes cannot. When it cannot, you get both, and the confirming test is the only thing that decides.
- The rest of the board. We see a component and the context you type. We do not see the fault. A correctly identified part fitted to a board whose real problem is upstream is a correct answer and a wasted afternoon.
- Stock, price and availability. We do not know whether a substitute is buyable this week. Nothing on a result page should be read as a purchasing recommendation.
- Anything you fit without testing. The confirming test is the field we work hardest on because it is the one that makes the rest safe. A likely candidate is a hypothesis with a measurement attached, and skipping the measurement throws away the part of this that is actually reliable.
The honest summary: this tool is at its best on a marked discrete or small IC where you can trace two or three pins, and at its worst on an unmarked passive on a board you cannot probe.
Try it on something you already know
The best way to judge this is to run a part you have already identified and see whether the reasoning matches yours. 3 a month are free, so it costs you a photo. Identify a part.