TL;DR: A parts search can give an exact, in-scope catalog identifier an explicit identity result before blending discovery rankings. Preserve the raw input, define accepted aliases with the catalog owner, and show ambiguity when multiple manufacturers use the same text. Evaluate exact lookup and product discovery by their different outcomes.
At an invented parts counter, a technician places a damaged fitting on a work mat. The label reads GT-401B. The apprentice types the code. The first search result is a family of fittings with a related description and an attractive picture. A more modest listing five places down contains the exact identifier. The apprentice clicks the first result and starts comparing thread details that should never have been a question.
The challenge for Linden Supply is to recognize both ways a customer approaches this counter. One person wants the identity carried by a particular label. Another describes an installation and needs to discover a compatible part. The box that accepts text can serve both. Its relevance ordering alone cannot state the catalog relationship between items.
The following conversations are reconstructed examples, and all records, ranks and counts are fictional. They give a product owner a way to choose where to apply identifier policy, how to scope matches, and what to require of a search migration.
The printed label
Technician: “That is GT-401B. I need the manufacturer page for this exact item.”
Apprentice: “Search says these fittings are close. It brought the most common one to the top.”
Catalog steward: “A related product family is a useful suggestion once we know whether we carry this item. Read the label record first.”
The steward looks up the explicit manufacturer's part number in a separately maintained field. The stored value is attached to a manufacturer namespace, a stable internal catalog row, and a source record from the manufacturer. A description field contains words an index might split, stem, or interpret. The identifier field has a narrower job. It should preserve meaningful distinctions between labels such as GT-401B, GT-401-B, and GT-4018 until a catalog authority confirms equivalence.
A field mapping lets the search system preserve this distinction. Elasticsearch, for example, provides a keyword mapping for structured values such as identifiers and distinguishes exact value retrieval from analyzed text search. The mapping alone does not decide which strings are aliases, or whether a match belongs to the intended manufacturer or customer catalog. Those relationships come from the application data. Elasticsearch keyword field type.
Imagine the actual printed label has a trailing blank where a scanner ended the reading. The identifier policy may recognize that scanner's framing convention and preserve the raw read separately from the candidate sought. The team tests the transformation against a set of real labels and intentionally similar examples. Removing all punctuation across every manufacturer risks collapsing two distinct products. An older import may have done just that. Its canonical field can be repaired only with catalog evidence.
The exact lookup result can include the name, original number, source, lifecycle state and compatible actions. In this conversation the stored part is discontinued. A result heading saying the item was found gives the technician confidence that the code was recognized. A separate status message says ordering has ended and links to the officially recorded replacement if one exists. Silently sending the technician to the replacement leaves them uncertain whether search recognized the code at all.
A claim about replacement is stronger than semantic similarity. A source entry listing a successor may preserve compatibility conditions, such as requiring a different gasket. The search system can display those conditions and the effective date. A second item with a similar description can remain in discovery, with its weaker relationship labelled by interface text. The customer who has a maintenance procedure can now verify it against the right item.
Two manufacturers use the same characters
Customer: “The label only says 121-A. It came out of an older system.”
Counter clerk: “Both products in front of me carry that number. Which family did it come from?”
Customer: “I have the equipment plate but the manufacturer's mark wore away.”
This time exact lookup returns two authorized catalog records, each in a separate namespace. Staying exact requires preserving both. A system that chooses the popular record as an invisible tie-breaker makes a claim about identity without evidence. An ambiguity view can show the two product families, known dimensions and whether documentary evidence would distinguish them. The clerk can continue discovery from the available context.
The matching policy specifies identifiers' types alongside their values. An internal SKU may be unique across Linden's whole catalog. A supplier ordering code might be valid only within a distributor account. A manufacturer number can reappear in another manufacturer's catalog. A scanned barcode has its own format and check rules. The UI may guide a customer with a clear field hint. It can also run several eligible lookups, preserving which types matched.
The same input can have a valid interpretation in two fields. In Linden's example the local inventory bin uses 121-A and the manufacturer's part number also includes 121-A. The bin belongs to an internal operational screen. Customer search has no reason to retrieve a bin as a saleable part. The route uses the product surface and authorized scope to decide where matches belong. That decision should be logged explicitly enough that a support agent can explain it to the person asking.
Partial queries require a different behavior. As someone types 121, a drop-down may offer several completions. Those suggestions answer what strings begin with the visible prefix. They are suggestions even if a single row happens to be returned. After the customer selects a specific full identifier, the lookup has a definite input and can give the typed match status. Instrumentation distinguishes a search submit from a completion selection so analytics reflect which behavior occurred.
A region also changes catalog scope. If a distributor maintains different records for a manufacturer code in two territories, the applicable catalog can show the authorized product and information available there. A search widget receiving the country as a stale client preference can puzzle someone who recently moved. Linden can display the active catalog and allow a controlled change. Internal reports keep query scope long enough to distinguish genuine misses from lookups that should never have returned the other territory's item.
The clerk walks away from the label
The third customer says they need a fitting that can be removed with gloved hands. They mention the line's physical clearance and the rough size, but provide no exact code. Good discovery here must bridge ordinary language and catalog descriptions. A lexical ranker can match measurement text and familiar product terms. A learned representation can retrieve differently worded descriptions. Filters for material, clearance and specification still need data aligned to the relevant facts.
An engineer combines ranks with reciprocal rank fusion. The published method adds contributions of the form one divided by a constant plus a document's rank in each included list. The method is a way to unite ranked lists, and its published empirical findings belong to the reported experiments. A typical illustrative calculation at Linden uses a constant of 60 and starts ranks at one. Reciprocal Rank Fusion paper.
For a mixed query, assume one ranking places a catalog match second and excludes a lookalike. Another ranking prefers the lookalike, and a third list also puts the lookalike first. Give the exact match only its rank-two contribution, 1 divided by 62, about 0.0161. The lookalike receives two rank-one contributions, totaling about 0.0328. Fusion gives the lookalike the larger score even though a dedicated identity procedure could have resolved the identifier. That arithmetic follows from the chosen rank inputs and says nothing about which fitting is interchangeable.
The team therefore treats a verified identity as a route with its own output. Discovery can still use multiple ranked lists. It can even show similar products next to a resolved identifier, with a visible boundary between the exact row and candidates. An application that suppresses all discovery whenever a label is present would disappoint someone asking whether a newer listing replaces their part. A clear supported relationship is more useful than one list pretending to represent all product meanings.
Consider a vendor page that contains a competitor's identifier in a compatibility note. A high lexical score is sensible because the term is present. If the note is backed by a curated compatibility claim, discovery can present it in the appropriate group. An exact identity result requires the value in the catalog record's applicable identifier field. A reader evaluating a purchase can distinguish a product that the text mentions from the product that owns the label.
The ranked lists may be restricted before or after blending. A filter applied after a small candidate window can leave very few eligible results despite there being many in-scope items further down each list. An exact lookup should run scoped eligibility on the identity field. Discovery should observe where filters act and how each candidate window is filled. That makes failures from scope, rank truncation and data freshness independently testable.
The steel drawer carries three reversible labels
The person maintaining the catalog separates the part's identifier from identifiers used on its packaging. A package holds ten identical items and carries its own ordering value. A supplier may stamp the item with the manufacturer identifier and file an offer under the package code. Exact lookup of the package should produce the unit displayed in the order flow. If the same result card shows price, the reader needs to know whether that price buys one part or the whole pack.
A change in packaging may be reversible from the search team's perspective and permanent from the buyer's. If an old pack ordering code was withdrawn and the individual piece remains available, the search result can preserve the discontinued package and show the alternative quantity. The catalog should not normalize both to the individual part and rely on a price fragment to distinguish them. A later warehouse report needs to know which saleable unit the user selected.
Time creates a third label. A manufacturer may publish an engineering revision and keep the original family code, adding a separate revision mark. The search input without the revision can identify the family while leaving a concrete version question unresolved. An authorized packaging alias can target a versioned record. A trade catalog that only contains the most recent version still needs its scope visible when someone maintains old equipment. Families, individual parts, packed units and revisions deserve independent stored relationships.
When old transcripts are imported into the new index, preserve the historical route. If a saved shopping list named a family code and recorded a specific accepted item, a migration can retain the original choice and link present availability separately. Linking to the current highest ranked result would mutate a person's past resolution. An inventory website can use unique internal addresses for saved items and mark searches as reproducible queries with changing results. The two link types should be visibly different in support investigations.
A receipt for a query the system could not settle
A customer enters GT-4O1B with the letter O in the middle. The catalog stores GT-401B with a zero. Linden knows this character pair is occasionally misread from worn labels. The team decides to show the latter as a suggested correction, with the original text still visible. An exact identifier badge belongs only to a supported alias or verified input. The operator can confirm the suggestion after checking the label or equipment documentation.
The lookup receipt keeps the input and its source, whether a scoped identity check had one match or several, aliases considered under approved rules, the used catalog version, and results shown. For a public search system the data collection design limits retention and access. Most investigations need a test identifier and a route outcome. A real raw query can incidentally include order or customer information, so it need not be a permanent debugging asset.
If the suggested item is accepted, define what that acceptance means. Clicking on the suggestion signals interest. A customer reaching the installation instructions after confirming a plate can support identity resolution with higher confidence. A commercial purchase has its own signals and delays. Linden can sample completed searches with a customer-service team or compare known item targets in task tests. Treating every click as a correct retrieval makes an attractive wrong listing reinforce itself.
A stale listing creates another class of unresolved result. A catalog import may retire a code from active orders while keeping old search data. In a transition, the exact lookup might use a recently updated registry and discovery might still show the product as orderable. A coherent interface can carry source timestamps and give the current ordering owner authority. During an index rebuild, version the catalog acceptance test and compare both views on known retired and replacement pairs.
The team designs a fallback for genuine absence. It can state that the active scope has no verified identifier match, show approximate candidates with their own labels, and provide a route to the right distributor or support flow. The phrase no exact match applies to the selected catalog revision. When a customer asks about a product beyond that scope, the team retains the uncertainty rather than implying the part has no real identity.
What the weekly counter totals say
Linden sets up a blinded task set with eight hundred fictional queries whose intended outcomes were recorded before scoring. Three hundred are unambiguous exact identifiers in scope, sixty have multiple exact owners, and forty require an approved alias or show absence. The remaining four hundred describe a use or constraint and require discovery. Accuracy on the first group and success on the last group answer separate questions, even when an executive asks for a single overall trend.
In one candidate system 279 of the three hundred unambiguous tasks show the exact record immediately, an illustrative 93 percent. Of the remaining twenty-one, twelve have a replicated catalog record that the system shows ambiguously, six match after a user correction, and three are missing from the current index. A second system reaches 290 immediate results, about 96.7 percent, after an identity routing change. A log of changes shows that nine records resolved through the new route and two after the catalog team repaired missing values. The improvement belongs to both workstreams and should be described accordingly.
In the ambiguous group, forty-five of sixty tasks show all valid owners without silently privileging one. Eight display a singleton because old feeds merged two records. Seven fail to include one authorized record in the candidate window. Any reported rate should name the rule used. Showing exactly one record in these cases is usually the wrong behavior. A global top-one relevance score could celebrate it anyway.
The discovery group needs graded relevance because several items may satisfy the requested use. The team can evaluate the shortlist and ask a human to validate compatibility evidence. Consider 400 cases, of which 240 yield at least one acceptable first-page result. Baseline and candidate differences should be paired on the same tasks to locate which classes gain or lose. Averages can hide a utility loss on very constrained searches when the route works well on ordinary fittings.
The exact and discovery groups also need clock measurements. Immediate identity responses can become slower if the application waits for every discovery channel. The product team can display a resolved exact result as soon as its authoritative branch returns, then load suggestions with an honest status. If orderability depends on a slow supplier response, the interface can show that state without changing the identity result. The latency target belongs to the interaction stage the user can act on.
Before release, Linden samples rare lookups by manufacturing family and region. Aggregate exact accuracy can depend on a dominant vendor's excellent data. Limitations in a smaller catalog remain visible when the test reports by scope. A repair to one identifier parser also deserves a regression set of deliberate neighbors, including letter and digit differences, vendor punctuation, retired part codes and codes with multiple valid meanings.
Publish the catalog version with the identity decision
A data engineer stages corrections from two sources at different intervals. The warehouse switches the pack record as soon as a shelf is relabelled. The manufacturer catalog sends a daily modification list. During the gap, one system may know the new code and another may understand the old relationship. Reporting a match requires naming the authority used for that match. Taking the freshest field from every source can construct an item neither authority intended.
The index builder uses a declared source revision or a staged inventory fence. As each searchable snapshot becomes active, its alias and identifier fields refer back to compatible catalog records. A dry evaluation can compare how each known query resolves under the prior and staged editions. The catalog owner reviews newly ambiguous and newly absent cases before customers see them. After activation, a server can include an index revision in its diagnostics so a retailer can reproduce disagreements.
Sometimes the right sales state updates faster than the search catalog. A returned identity should continue naming the same part even when the order page is waiting for a fresh availability result. If the availability check returns a version discrepancy, the UI can present the situation and preserve the link to the authoritative identity record. This preserves the operator's work while the system resolves exactly which pack it can fulfill.
The case for a taller result card
The apprentice finally returns to the first customer. The result card names GT-401B, shows the manufacturer, states the retirement date, and identifies the documented successor and its installation condition. The old fitting stays available as a record, consistent with the technician's original lookup. A separate path offers other fittings by physical criteria.
The team keeps this record as an acceptance example for every search change. In the next import a new alias arrives with the same characters. The catalog authority decides whether it is a duplicate part, an alternate label or a genuinely separate row. The service then updates identity behavior based on a traceable decision. A ranker can change the order of similar items without rewriting that history.
That distinction survives a platform migration. If a new search provider changes punctuation, ranking or candidate limits, the catalog still owns identifier equivalence. Migrating the retained exact test set and scope decisions matters more than matching old relevance scores. The review panel can compare difference counts at each stage, and the counter clerk can name the unresolved cause when a familiar label moves.
A search rollout earns the technician's trust when it can say which relationship it found and what work remains. An exact lookup establishes an item record, an equivalence decision establishes an alias, and a product specialist establishes compatibility under stated conditions. Discovery connects evidence for consideration. All of those services can fit in one result card while retaining their different meanings.