Module 6: Beyond Similarity (Bonus) on Qdrant - Vector Search Engine

Module Overview

info@qdrant.tech (Andrey Vasnetsov) — Mon, 01 Jan 0001 00:00:00 +0000

Module 6

Beyond Similarity

Modules 1 through 5 showed you how to design and build a complete retrieval pipeline. This bonus module covers the next layer: measuring and improving its results.

Overview

You’ll match common search problems to Qdrant features: score boosting and reranking for order, Maximal Marginal Relevance (MMR) for variety, grouping for one slot per document, and the Recommendation and Discovery APIs for searches from examples instead of text. You’ll set up a way to measure relevance first, and pick up the trap that comes with each feature along the way.

Find Your Problem

info@qdrant.tech (Andrey Vasnetsov) — Mon, 01 Jan 0001 00:00:00 +0000

Module 6

Find Your Problem

Each row pairs a problem with the feature that addresses it, and names the stage of a query where it runs.

Problem	Feature	Stage
The right documents come back in the wrong order	Score boosting with a formula query	Rescore
The order should account for recency or distance	Decay functions inside the formula	Rescore
The accurate model is too slow to run over the whole collection	Multi-stage query	Rescore
The top results are near-identical	Maximal Marginal Relevance	Select
One document fills the page with its own chunks	Grouping	Select
A better model, or user clicks, disagree with retrieval	Relevance feedback	Next query
There is no query text, only examples of good and bad	Recommendation and Discovery APIs	Replaces the query

Measure First

info@qdrant.tech (Andrey Vasnetsov) — Mon, 01 Jan 0001 00:00:00 +0000

Module 6

Measure First

Ranking changes are hard to judge by eye, because a worse results page still looks like a list of plausible documents. Measure what you have before you change anything.

A golden set pairs queries with the documents that should come back for them. It turns a ranking change into a number.

Sample query and click pairs from your logs, or have someone who knows the domain write 20 or 30 queries with the answers they expect. Measuring Retrieval Relevance covers both and computes the metrics with the Python library ranx.

Ranking: Score Boosting and Reranking

info@qdrant.tech (Andrey Vasnetsov) — Mon, 01 Jan 0001 00:00:00 +0000

Module 6

Ranking: Score Boosting and Reranking

Two ways to change the order when the right documents are already coming back.

Score Boosting

Similarity is not always the final ranking signal. A result may be relevant, but you may still want to prefer an exact title match, a nearby store, or a recent article. A formula query lets you rescore the candidates returned by retrieval, combining their similarity score with payload values and conditions you define.

Diversity: Maximal Marginal Relevance

info@qdrant.tech (Andrey Vasnetsov) — Mon, 01 Jan 0001 00:00:00 +0000

Module 6

Diversity: Maximal Marginal Relevance

Maximal Marginal Relevance (MMR) picks results one at a time, preferring candidates that match the query and differ from what it has already picked. In Qdrant it is a parameter on a nearest neighbors query, and diversity sets how much relevance it trades for variety.

The trap is candidates_limit. It defaults to the query’s limit, which leaves MMR nothing spare to choose from, so all it can do is reorder the results it was already given. This is the most common reason MMR looks like it did nothing.

Grouping: One Slot per Document

info@qdrant.tech (Andrey Vasnetsov) — Mon, 01 Jan 0001 00:00:00 +0000

Module 6

Grouping: One Slot per Document

Chunking creates the neighbor problem: one long document becomes many points, and a strong match on it can fill the whole first page with its own chunks. query_points_groups groups results by a payload field and returns a set number of groups, so one document takes one slot.

group_size caps how many chunks come back inside each group, and with_lookup attaches a parent record from another collection.

Searching From Examples and Feedback

info@qdrant.tech (Andrey Vasnetsov) — Mon, 01 Jan 0001 00:00:00 +0000

Module 6

Searching From Examples and Feedback

Some searches have no query text. A reader clicks “more like this”, or an analyst has three documents that are right and two that are wrong and no words for what separates them.

The Recommendation API searches from positive and negative examples. The Discovery API takes context pairs, where each pair names one region of the vector space to move toward and one to move away from.

Inspecting a Collection

info@qdrant.tech (Andrey Vasnetsov) — Mon, 01 Jan 0001 00:00:00 +0000

Module 6

Inspecting a Collection

Two features for checking what you actually ingested.

Feature	What it does
Facet counts	Counts how many points hold each value of a payload field, which also shows how selective a filter would be.
Random sampling	Returns a random subset of a collection, for spot-checking ingested data. For a subset that repeats across queries, such as an evaluation set, use the slice filter condition instead.

Knowledge Check

info@qdrant.tech (Andrey Vasnetsov) — Mon, 01 Jan 0001 00:00:00 +0000

Module 6

Knowledge Check

You turn MMR on and your Recall@10 drops. Is MMR broken?

No. MMR spends result slots on documents further from the query, so a measure that only counts relevance goes down while the page stops repeating itself. Check what those slots held before: if they were eight versions of one story, the drop bought something your metric cannot see. If they were eight distinct relevant documents, lower diversity.