Give a known part code a direct route
If a customer supplies a complete part number, look it up in the catalogue’s identifier field first. Keep the original characters alongside any cleaned-up version. A space might be optional in one supplier’s code, while a suffix might distinguish two versions that cannot be interchanged. Get the catalogue owner to specify which changes preserve the meaning.
The field you search matters. Elastic documents how its standard text analyser can strip punctuation, split a string into words and change case. Its term query can find an exact indexed value, such as a product ID. For a small catalogue, an ordinary database lookup against a unique part-number field may already do the job. Try that with a pair of easily confused codes before adding anything more elaborate.
An exact hit establishes which record the code identifies. If the customer also describes a different size or fitting, show the discrepancy to the desk. If the code is missing from the catalogue, ask the customer to check the label before offering a similar-looking code as a substitute.
Use the customer’s words to build a shortlist
Keyword search is a sensible starting point for descriptions that contain distinctive terms. Ask the parts desk for the names customers use most often. Adding “bottom wheel” as an alternative name for “lower roller” may fix a common missed search at little ongoing cost. Check whether that name is also used for a different assembly.
Vector search turns descriptions into numerical representations so that phrases with related meanings can match even when their words differ. Hybrid search runs a text search and a vector search and combines the results. Microsoft’s Azure AI Search documentation describes both approaches and singles out product codes as a case where keyword search can be particularly useful.
One way to combine the lists is reciprocal rank fusion. It uses each item’s position in the lists. An assembly near the top of both can move ahead of one that appears in only one list. Its final position tells the adviser which record to inspect first. The dimensions, revision and fit still have to come from the catalogue.
Illustrative example — two rollers, one missing measurement
Suppose a parts merchant stocks two lower rollers for the same sliding door series. Both catalogue entries say “lower roller assembly”. One fits a 32 mm track and the other a 38 mm track. The customer’s photo shows where the wheel sits, but gives no scale. These example codes and dimensions are fictional.
The desk can now ask one short question. “Can you confirm the track width, or send the part number from the old assembly?” Showing the adviser both widths makes that question possible. Showing the customer a quotation for the first search result would hide the reason to ask.
| What the customer supplies | What the desk can establish | Next action |
|---|---|---|
| “The little wheel under the door” | Both rollers are candidates | Ask for the track width or part code. |
| “It runs in the 38 mm track” | Only the 38 mm record fits that width in this pair | Confirm the door series before quoting. |
| “LR-32, for a 38 mm track” | The code identifies the 32 mm version and conflicts with the stated width | Hold the quote and resolve the discrepancy. |
Test what reaches the adviser
Take a sample of past enquiries and hide the final order from the search system. Ask an experienced adviser to mark the acceptable catalogue records and any detail they needed from the customer. Keep enquiries with unknown codes and insufficient information in the sample. Otherwise a system that always suggests a part can look deceptively good.
Run the same enquiries against keyword search and the proposed hybrid search on the same catalogue snapshot. For enquiries with a known answer, count how often an acceptable item appears in the first three results an adviser sees. Keep a separate count of wrong variants offered as definite matches. For ambiguous enquiries, check whether the system surfaces the missing detail and lets the adviser ask for it.
Read the misses by enquiry type. If both searches miss a part because its catalogue entry omits the track width, improve the entry. If descriptions keep missing the right assembly and hybrid search finds it, time how long advisers spend checking the extra candidates. That gives the team a way to weigh better coverage against checking time and the work of maintaining a second index.
Try one product family at the parts desk
Choose a family with similar variants and a named catalogue owner. For the first trial, show the matched code, description and deciding specification beside each suggested item. Let an adviser record “none of these” and the question they asked. Keep quote creation in the existing workflow while you learn which searches staff can rely on.
At the end of the trial, pick one recurring miss and trace it from the customer’s wording to the adviser’s decision. A missing synonym, an incomplete specification and a poor ranking each call for a different fix. Have another adviser try the revised search on fresh enquiries before extending it to the next product family.