When a user cannot find a product, a document, or a folder, the problem is not always the algorithm. The data may be incomplete, the filters incomprehensible, the permissions incorrectly applied, or the vocabulary different from that of the content. A useful search is a complete product that links indexing, relevance, interface, and measurement.
Vector search provides the ability to bring formulations closer together. It does not replace the precision of full text, exact references, or business filters. In many cases, a hybrid combination is the best solution, provided it can be evaluated.
Start with the research tasks
The queries express different intentions:
- to reach an item by its name or reference;- explore a category;- filter according to attributes;- find content that answers a question;- compare options;- pick up a recent file;- discover associated information.
The engine must recognize these families and the interface must support them. A single bar can offer search, filters, suggestions, and history without handling all intents in the same way.
Data quality precedes the algorithm
The index must have reliable titles, descriptions, taxonomies, identifiers, dates, languages, statuses, and permissions. Duplicates, empty content, and inconsistent units degrade relevance.
A document indexing pipeline:
- sources of truth;- transformations;- searchable fields;- filterable fields;- linguistic analysis;- update frequency;- deletion;- errors and resumption.
Freshness must be visible when indexing is asynchronous.
The full text remains essential
Lexical engines analyze words, their forms, positions, and frequencies. They are effective for proper nouns, references, exact expressions, and well-structured content.
The settings concern:
- weighting of the fields;- analysis by language;- synonyms;- fault tolerance;- closeness of terms;- popularity or freshness;- controlled business rules.
A product reference should not be corrected as a common mistake. Technical fields may use a specific parser.
Filters and facets give control
In a catalog or a business application, the user often knows certain constraints: status, date, site, category, certification, or owner. Facets reduce the search space and explain the structure of the results.
Values must be normalized, understandable, and accompanied by consistent accounts. Active filters are visible, removable, and shareable in the URL when this does not reveal sensitive information.
What vector search brings
An embeddings model represents queries and content in the form of vectors. Proximity can find a paraphrase or a concept even without common words. It helps with natural questions, long content, multilingualism, and heterogeneous vocabularies.
Its limitations are significant:
- exact references sometimes misfiled;- sensitivity to the model and the segmentation;- difficulty explaining a score;- indexing cost;- model update;- need to filter the rights correctly.
Vector search must be tested on the real corpus.
Compare the four research bricks
Full text
Precise, fast, and explainable for terms and references.
Structured filters
Essential for business constraints and exploration.
Vector search
Useful for semantic similarity, questions, and varied formulations.
Reranking
A more expensive model reorders a small set of candidates. It can improve the top of the results, but adds latency and dependency.
Pipeline of a hybrid search from indexing to fusion, reranking, and evaluation.
Build a hybrid search
A hybrid search executes several strategies and then merges their candidates. A reciprocal rank method can combine positions without directly comparing scores of different nature. Rules give priority to an exact match or an explicit filter.
The pipeline can be:
- analyze the request;2. apply security and context filters;3. retrieve lexical candidates;4. retrieve vector candidates;5. merge;6. apply business rules;7. rerank a subset;8. generate excerpts and explanations.
Each step has metrics and a degraded mode.
Respect the rights
An internal search must never reveal the existence of a document prohibited by its title, excerpt, or a facet count. Permissions must apply in the query or in a segmented index.
Filtering only after retrieving ten results can produce an empty page even though allowed results exist further on. The strategy must preserve both security and recall.
Legal changes must propagate quickly and caches must be segmented.
Index long documents
A document can be divided into sections or passages with a link to the parent. The division must respect titles, paragraphs, tables, and metadata. Pieces that are too small lose context; too large ones dilute relevance.
For documentary research, the result can display the relevant passage while opening the document at the appropriate location. The version and rights remain those of the source document.
Multilingual
Lexical analyzers and vector models must be compatible with languages. Can a French query retrieve an English document? The answer depends on the need and must be explicitly tested.
Synonyms are managed by language and profession. Automatic translation of queries can help, but it must preserve proper names and references.
Suggestions and correction
Autocomplete can suggest queries, categories, or known objects. It must respect rights and avoid revealing private data. The suggestions are fast, limited, and distinguished from the results.
Spell correction should not override the query. Displaying 'results for' with the option to return to the original expression is safer, especially for codes.
Evaluate before paying
An evaluation game contains representative queries, their intent, and judgments on the results. It covers successes, exact references, ambiguous queries, languages, filters, and cases without answers.
Technical metrics can include:
- Recall@k: presence of relevant results among the top candidates;- MRR: position of the first relevant result;- nDCG: quality of the ranking when multiple relevance levels exist.
They are complemented by useful click-through rate, reformulations, absence of result, task time, and business conversion.
Analyze queries with no results
An absence can mean missing content, unknown vocabulary, error, overly restrictive filter, or insufficient rights. The report must group the requests by themes and avoid recording unnecessary sensitive data.
Improvements can concern the content, synonyms, taxonomy, or the interface, not just the engine.
Explain the ranking
A simple explanation reinforces confidence: title match, exact reference, filter, semantic proximity, or popularity. Sponsored or commercial rules are identified.
Teams must be able to inspect a request, its candidates, and the scores in a secure administration tool.
Exploit and develop
The engine monitors latency, errors, indexing delay, size, slow queries, and metric drift. A new version of the vector model requires reindexing, comparison, and the possibility of rollback.
Relevance changes are deployed through experimentation or on a sample, with an analysis by intent so that an overall improvement does not hide a regression on exact matches.
Relevance comes from the complete system
The best engine does not compensate for poor data or inconsistent rights. Successful research combines structure, lexicon, semantics, interface, and evaluation loop.
Partitech can audit an existing search, design indexing, develop a full-text, vector, or hybrid engine, and set up an evaluation set. The goal is to make the results useful, safe, and measurable, not simply more sophisticated.
Let's talk about your project
Design or improve your search engine with Partitech.Contact Partitech.