Despite ANN dominating production retrieval, Flat (brute force) search remains the right call in two specific situations. Name and briefly justify both.
Free · 2 AI evals / day
**Flat search** is brute force scan: compare the query vector to every vector in the corpus and return the K with the smallest distance. It is O(N), exact, and the simplest possible vector search. It is also the baseline that every ANN method is measured against.
In 2026 production, two situations still call for Flat.
**(1) Tiny corpora.** Below roughly 100k vectors, building an HNSW or IVF index costs more (in engineering time, memory overhead, and parameter tuning) than the brute force scan saves at query time. A 50k-vector corpus with 1024-dim float32 embeddings is 200 MB; a single CPU can scan that in 2-5ms. ANN cannot beat that meaningfully and adds operational complexity. Pinecone, Weaviate, and Qdrant all expose Flat (or 'FLAT' / 'BRUTE_FORCE') as an explicit per collection option for this regime.
**(2) Ground-truth generation for ANN tuning.** When you tune an HNSW M and ef_search, an IVF nlist and nprobe, or a PQ pq_m, you measure recall@K. Recall is defined as the fraction of the true top-K that your ANN method retrieves. The only way to know the true top-K is to run Flat on the same query set. Every recall number in every ANN benchmark traces back to a Flat baseline.
A third niche: regulated workloads. Legal discovery, certain financial-compliance contexts, and some healthcare use cases contractually require 'exact nearest neighbor' rather than 'approximately nearest neighbor.' Flat is the only method that satisfies the spec.
**Flat is brute force search:** compute distance to every vector, sort, return top-K. It is O(N) in corpus size and exact by construction.
**Situation 1: tiny corpora.** Below roughly 100k vectors at typical embedding dimensionality (768-1536), Flat returns in single digit milliseconds on a single CPU core. ANN at that scale adds build time (HNSW graph construction is non trivial), memory overhead (the graph itself plus index metadata), and parameter-tuning surface area, all to save microseconds the user will never feel. Pinecone exposes Flat via `metric=cosine, index_type=flat`. Weaviate has a Flat vector index type explicitly for small classes. Qdrant offers `on_disk=true` with no graph as the equivalent. The 2026 reality: any corpus that fits comfortably on one machine in raw form and serves under 100 QPS is a Flat candidate.
**Situation 2: ground truth for ANN tuning.** Every recall@K number in every ANN benchmark depends on Flat. Recall is defined as `|ANN_topK ∩ true_topK| / K`, and the only way to compute `true_topK` is to brute force the query set. When tuning HNSW M, ef_search; IVF nlist, nprobe; PQ pq_m; or any quantization scheme, the workflow is: (a) generate a representative query set, (b) run Flat on the corpus for each query to get the true top-K, (c) sweep the ANN parameter, (d) compute recall against the Flat baseline at each parameter setting. Without the Flat baseline, you are tuning blind.
**Situation 3 (niche but real): regulatory or contractual exactness.** Legal discovery (eDiscovery), certain financial-compliance workloads, and healthcare contexts that require deterministic similarity matching cannot use ANN because 'approximately' is not a defensible word in those contexts. Flat is the only method that satisfies an 'exact nearest neighbor' contract.
Quick reference
TL;DR
Two situations: (1) tiny corpora under ~100k vectors where ANN's build overhead exceeds the savings, and (2) generating the exact recall@K ground truth used to tune any ANN index.
Imagine you have a shoebox of 500 baseball cards and a friend hands you one card asking 'which of mine is closest to this?' The dumbest possible answer, flip through every single card, takes a few seconds and you're done. Building a fancy filing system would take longer than the answer itself. **That's the tiny-corpus case.** The other situation is when you're tuning the fancy filing system: you need to know the right answer to grade it on, and the only way to know the right answer is to flip through every card. Flat is slow but it never lies, which is exactly what you need when you're measuring how often a fast method tells the truth.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.
Corpus exceeds 1M vectors and the latency SLO is sub-100ms Flat's O(N) cost becomes prohibitive; HNSW or IVF is the right move.
QPS exceeds a few hundred on a corpus over 100k vectors Each Flat query saturates a CPU core for milliseconds; the throughput collapses. ANN's sub linear cost is needed.
The corpus will clearly grow into HNSW territory in the next quarter Avoid the migration tax; build ANN from the start if the trajectory is known.
Quick scan · 2–4 min
Everything important, quickly.
3 min: name both situations (tiny corpora + recall ground truth), justify each with concrete numbers, cite a vendor that exposes Flat, mention the regulated-context niche.
Side-by-side
Property
Flat
HNSW
IVF
Recall
1.0 (exact)
Tunable, typically 0.90-0.99
Tunable, typically 0.85-0.97
Build time
Zero
Minutes to hours
Minutes (k-means)
Memory overhead
Just the vectors
Vectors + graph edges
Vectors + centroids
Query latency at 100M vectors
Seconds-minutes
Sub-millisecond
Single-digit milliseconds
Query latency at 10k vectors
<1 ms
<1 ms (but build overhead)
<1 ms (but build overhead)
Use case in 2026
Small corpora + ground truth + regulated
Default production ANN
Large corpora + compression
Real-world examples
Real products, models, and research that use this idea.
1Pinecone exposes a Flat index type explicitly for tiny corpora and offline ground truth jobs.
2Weaviate ships a Flat vector index recommended for classes under 10k objects.
3Qdrant supports collections without an ANN index for the same regime.
4Every ANN benchmark on ANN-Benchmarks.com reports recall against a Flat baseline.
5Faiss ships `IndexFlatL2` and `IndexFlatIP` as the canonical ground truth indexes used in every production evaluation script.
Interviewer follow-ups
What an interviewer would ask next. Try answering before peeking at the approach.
QHow would you generate a Flat ground truth for a 100M-vector corpus when running Flat on the full corpus is infeasible?
Sign in or go Pro to see the approach for this follow-up.
QWhy does Faiss ship both IndexFlatL2 and IndexFlatIP separately rather than unifying them?
Sign in or go Pro to see the approach for this follow-up.
QWhen tuning HNSW, how often should you regenerate the Flat ground truth?
Sign in or go Pro to see the approach for this follow-up.
Don't say this
Red flags and common mistakes that signal junior thinking. Click to expand.
Most common mistake
Assuming Flat is always obsolete because ANN is faster. Flat is the only exact method, which means it is the only valid ground truth for measuring an ANN index's recall, and it is the cheapest option on small corpora.