Search for .NET
Better search in your .NET app. No server to run.
IndxSearchLib is the Indx engine as a NuGet package: typo-tolerant, ranked search over your own JSON, in your application’s own process, in under a millisecond.
Your first search
From JSON to results in four steps
No analyzers, tokenizers or language settings to choose: you say which fields are searched, and Indx does the rest. This is the getting started guide, in short.
- 1
Load your JSON
An array of objects, nested ones too. Init reads which fields there are.
using Indx.Api; var engine = new SearchEngine(); using var stream = File.OpenRead("products.json"); engine.Init(stream); - 2
Say what is searched
Nothing is searched until you say so. Whole words, and inside words; a weight for what matters most.
var name = engine.GetField("name")!; name.WordMatching = true; name.PatternMatching = PatternMatchingSetup.Default; name.Weight = 2.0f; engine.GetField("category")!.Facetable = true; engine.GetField("price")!.Filterable = true; - 3
Index it
Stream the same JSON again, and build the index. The engine is ready.
stream.Position = 0; engine.Load(stream); engine.Index(); - 4
Search
Typos, half-typed words and filters, in one query.
var cheap = engine.CreateRangeFilter("price", 0, 50, out _); var result = engine.Search(new Query("wirless headphnes", 20) { Filter = cheap });
Better results
Finds what people mean, out of the box
Up to three typos in a word, and pattern matching past that. Every form of a word from one search. The title that says it before the description that mentions it. All without a line of configuration.
- 1HonningmelonWhole word
- 2HoninmelonTypo
- 3Honning melonSplit
- 4Melon honningSplit, turned
And, when the query allows part of it
- 5HonningPartial
- 6MelonPartial
Typo tolerance
Up to three typos in a word, counted by edit distance. Past that, pattern matching keeps guessing, so even a badly mangled word can still find its record.
Every form in one query
Whole words, the start of a word, misspelled words, and words written together or apart, in the same query. No second search, and no separate autocomplete.
Accents, one way
Search for jalapeno and you get jalapeno and jalapeño alike. Type the ñ on purpose, and the records spelled that way come first. On by default.
Search as you type
A full ranked search on every keystroke, with the word being typed read as the start of a longer one. No separate suggest endpoint.
Built-in relevancy ranking
Search for titanic, and Titanic comes before A Night to Remember, which only mentions it in its description. A word in a short field like a title counts for more than the same word in a long text, without weights or tuning.
Field weights
Say that a title counts for more than a description, for every search or for one.
Hybrid search
Bring your own embeddings, and blend vector and text scores in one ranked list.
Any language
No tokenizer, stemmer or stop words to set up per language. Compound words are split as the index is built. Irregular forms, like mouse and mice, take a synonym.
Synonyms
Teach it the words your visitors use for yours, one way or both. They apply on the next search, with nothing re-indexed.
Your JSON as it is
Nested objects by dot notation, and arrays matched on any of their elements.
What a catalogue needs
Filters, facets and sorting, kept current
Counted over the whole dataset, so the numbers and the list agree; and a change is searchable the moment you make it.
Filters
Values, ranges and true or false, combined with AND, OR and NOT, over the whole dataset and not just the page.
Facets
A count for every value, and the lowest and highest of a number, that always agree with the count beside the list.
Sorting
A field orders equally good results, so the best match stays first. With no search text, the field is the order.
Under a millisecond
Embedded, a query runs in your own process in under a millisecond, with no network round trip and no cold start.
Real-time updates
A change is searchable at once: no queue, no rebuild. A whole dataset can be replaced with no downtime.
Hibernation
A dataset can give its memory back and keep its documents, and be woken again when it is needed.
Questions
Before you add it
- Do I need a server?
- No. IndxSearchLib runs in your application’s own process, with no network hop and nothing else to deploy. If several applications, or ones not written in .NET, should share an index, run the Indx server instead: the same engine behind an HTTP API.
- Which .NET does it need?
- .NET 10. Version 6.0 is in beta on NuGet, so add it with --prerelease for now.
- Is it free?
- Every search feature is in the free version. Without a licence file the library indexes up to 100,000 documents; a free licence from license.indx.co removes the limit. Pro and Enterprise add support.
- How do I keep the index current?
- Add, update and remove documents as they change; a change is searchable at once, with no queue and no rebuild.
- Does it do vector search?
- Yes, with your own embeddings: Indx stores and searches vectors, and blends them with the text score in one ranked list.
- Can my coding agent set it up?
- Yes. The Indx skill teaches Claude Code and other agents the library’s API, the field roles and the pitfalls: npx skills add indxSearch/skill-indx-search.