Three ways to search for a movie

Matches terms in the title and original_title fields, both of type text. Elasticsearch analyzes the query text into tokens and compares them with the indexed tokens of each title.

Available settings

Query type: match

Searches a single field, which you choose: title or original_title.

Query type: multi-match

Searches both fields at once with the best_fields type: each field is scored separately and the movie gets the score of the best-matching field.

Operator: OR

Any one of the query words is enough. This is Elasticsearch's default behavior.

Operator: AND

All the words must appear. With multi-match and best_fields, they must all be in the same field: they can't be split between title and original_title.

Documentation Match query Multi-match query Similarity (BM25)

Searches by meaning in each movie's plot (the plot field). Your query text is turned into a vector with the jina-embeddings-v5-omni-small model, and Elasticsearch returns the movies whose plot vectors are closest, using the knn retriever.

The three plot representations

This mode is meant to test how the number of vector dimensions affects result precision. All three fields come from the same model: its native dimension is 1024, and it supports Matryoshka truncation to 256 and 512.

A fair comparison. All three fields use the same index settings: index_options of type hnsw with unquantized float vectors, the same m (16) and ef_construction (100), which are Elasticsearch's defaults, and cosine similarity. That way, the only difference between them is the number of dimensions.

Available settings

Vector k-NN

The field to search: plot_256, plot_512 or plot_1024. More dimensions usually capture more nuances of meaning, at the cost of more memory and more expensive comparisons.

k

Maximum number of results the vector search returns, after merging the candidates from all shards; they are split into pages of 25, 50 or 100. Must be less than or equal to num_candidates. Initial value: 100.

num_candidates

Candidates evaluated on each shard before keeping the best k. Raising it tends to improve precision, at the cost of longer response times. Maximum 10,000. Initial value: 200.

Implementation note: for retrieval, the model expects the Query: prefix on queries. Plots are vectorized with input_type ingest and no prefix.

Documentation kNN retriever Approximate kNN Dense vector jina-embeddings-v5-omni-small

Combines the two modes above with the same settings. The rrf retriever wraps a standard retriever, with the match or multi_match query from the full-text component, and a knn retriever from the vector component.

1. 1 · Full-text component List sorted by BM25 over title and/or original_title.
2. 2 · Vector component List sorted by k-NN proximity on the chosen plot vector.
3. 3 · RRF fusion A single final list, sorted by each movie's position in both lists.

Fusion parameters

The demo doesn't expose them in the UI: rank_constant uses Elasticsearch's default and rank_window_size is set by the application.

rank_constant = 60

The higher it is, the more weight movies further down each list get.

rank_window_size = k

How many results from each component go into the fusion, and how many combined results you can page through. The application sets it equal to k (100 by default), so both components contribute the same number and pages of 25, 50 or 100 are never empty.

Documentation RRF retriever

Results are shown in pages, with the score that matches each mode, and can be narrowed with filters computed from the results themselves.

Pagination

25 · 50 · 100 per page

You choose the page size and move with Previous, Next or a page number. Changing the size or applying filters takes you back to page 1.

Full-text

You can page through all matches with from and size, up to the 10,000 limit set by index.max_result_window.

Vector

There are at most k results (100 by default), split into pages.

Hybrid

There are at most rank_window_size combined results. The application sets it equal to k, so from + size never exceeds the window and no page is empty.

Score of each result

Full-text

BM25 score. It has no fixed scale: it's only useful for comparing results within the same search.

Vector

k-NN similarity. With cosine, Elasticsearch returns (1 + cosine) / 2, a value between 0 and 1.

Hybrid

The BM25 score and the k-NN similarity, each with the movie's position in its list, plus the RRF score that sets the final order.

Filters

Fields

adult, genres, release_date (from–to range), spoken_languages and rating.

Combination

Across different filters, all must match; within a single filter, any one of the checked values is enough. Changes take effect when you press APPLY; the “From” date can't be later than “To”.

Dynamic values

Each filter shows only the values present in the current results, with how many results have each one, and they're recalculated whenever the results change. Values you've already checked stay visible so you can uncheck them.

In Elasticsearch

Full-text: filters go in the query's filter context, which doesn't affect the BM25 score, and the values come from terms aggregations. Vector: filters go in the filter of the knn retriever, which applies them during the search, so it returns k movies that match them. Hybrid: they go in the filter of the rrf retriever, which applies to both searches.

Values in Vector and Hybrid

The application computes them from the returned results (up to k). With RRF, aggregations would include every match from each search, even movies that don't appear in the combined list.

Documentation Pagination Terms aggregation Filtered kNN search RRF retriever

You can add movies one at a time or import a file. A file can be loaded in two modes: append, which adds the movies to the catalog, or bulk, which replaces it. In both, the whole file is validated before anything is written.

Ways to load data

Single record

A form with the mapping fields. id, title and plot are required.

.json file

The index export (an object with documents and, optionally, mapping and document_count), an array of movies, or NDJSON.

.csv file

UTF-8 text with or without BOM, comma-separated, with a header. A CSV without a header is rejected. Lists and nested objects go as JSON inside the cell.

Validation

Before uploading

Only .json or .csv, no double extensions, up to 25 MB, with UTF-8 text content. The server repeats these checks.

Against the mapping

The whole file is checked: every field must exist in the mapping and have its type (release_date as yyyy-MM-dd, adult as a boolean, keywords and belongs_to_collection as nested objects). id can't be repeated.

If something doesn't match

The load stops and no records are imported. The page shows each error with its record, id, field, received value and what the mapping expects, and lets you download the full error report.

Warnings

Anything that doesn't break the mapping doesn't stop the load: for example, empty values inside a list are skipped and empty strings are stored without a value.

Import

Vectors

The three vectors are always generated from plot with the Elasticsearch inference service (jina-embeddings-v5-omni-small, input_type ingest and no prefix). If the file includes plot_256, plot_512 or plot_1024, they're ignored.

Append mode (default)

The movies in the file are added to the current index through the alias, in batches of 100 and with no temporary index. If an id already exists, the movie is replaced. If a batch fails or the load is cancelled, whatever was already written stays and the page tells you how many movies were added.

Bulk mode

The movies in the file are written in batches of 100 to a new temporary index that holds only the movies in the file: the current catalog is not copied.

Publication (bulk)

When it finishes, the count is checked and the elastiflix-reloaded alias moves to the new index in a single atomic operation. The previous index is deleted.

If it fails (bulk)

If Elasticsearch rejects a batch or the check doesn't add up, the temporary index is deleted and the catalog stays exactly as it was.

Documentation Index aliases Dense vector

The catalog lives in elastiflix-reloaded, an alias that points to the current index: searches always read the published catalog. A bulk load is published all at once or discarded without affecting searches; an append load is added to the current index batch by batch. It has 6,899 movies. These are all its properties, the ones in elastiflix-reloaded-mapping.json, grouped as in the «Load data» form. The «Used in» column says where the demo uses them: the card is each search result and the detail, each movie page.

The last two columns say what a file must include to be loaded, in append or bulk mode (the rules are the same): in a .json, every movie has id, title and plot with a value; in a .csv, the header has those three columns and every row, their value. The other fields are optional, but if present they must match their type, and a field or column that is not in these tables rejects the file.

Identification and titles

Plot and text

Rating, date and status

Metrics

Cast, production and media

Credits and terms of use

Credit for the movie information goes to TMDB and Wikipedia.

Movie data and poster images come from TMDB (The Movie Database).

This product uses the TMDB API but is not endorsed or certified by TMDB.

Este producto usa la API de TMDB, pero no está avalado ni certificado por TMDB.

Movie texts come from Wikipedia, whose content is published under the Creative Commons Attribution-ShareAlike 4.0 license.

Names, trademarks, posters, and texts belong to their respective owners. This demo is not affiliated with or endorsed by TMDB or Wikipedia.