阅读中文版本

Build a Vector Embedding Store

Vector Embedding Store — a complete walkthrough of vector similarity search, covering hnsw index, distance metric, insert pipeline, and testing.

vector_embedding_store/
├── src/index.ts
├── src/hnsw-index.ts
├── src/distance-metric.ts
├── tests/vector_embedding_store.test.ts
└── package.json

// HNSW index: vector similarity search core
export function hnswindex(input: HnswindexInput): HnswindexResult {
  // Build the hnsw index that anchors vector similarity search and defines the core data shape.
  return processHnswindex(input);
}

// Distance metric — handles vector similarity search Add the distance metric so
export class Distancemetric {
  constructor(private opts: DistancemetricOptions) {}
  async run(ctx: Context): Promise<DistancemetricOutput> {
    return this.execute(ctx);
  }
}

// Insert pipeline: vector similarity search stage
export const insertpipeline = (next: Next) => async (req: Request) => {
  const out = await transform(req);
  return next(out);
};

// Search reranking for vector similarity search
export async function searchreranking(stream: AsyncIterable<Chunk>): Promise<Aggregate> {
  let acc = init();
  for await (const c of stream) acc = merge(acc, c);
  return acc;
}

// Metadata filter: vector similarity search finalization
export function metadatafilter(state: State): Result {
  validate(state);
  return finalize(state);
}

Project overview

This project builds a vector similarity search system whose reason for existing is finding the k nearest neighbors over millions of floating-point vectors without scanning every one. The temptation in vector similarity search is to bolt the interesting part onto a working prototype and discover the hard cases in production; this build inverts that order and treats a high insert rate that degrades the graph before a background compaction can re-link nodes as the first thing designed, not the last.

You will implement HNSW index, Distance metric, Insert pipeline, Search reranking, and Metadata filter, each as a module with one responsibility and a test that proves it in isolation. The vector similarity search pipeline is then the wiring that composes them, and the integration test is where the composition is shown to survive a high insert rate that degrades the graph before a background compaction can re-link nodes.

The walkthrough is written for a developer who has used a vector similarity search system but has not had to make one honest. It favors explicit data shapes over clever abstractions, and it names the failure each design choice prevents rather than presenting the choice as a style preference.

Each phase ends with something observable — a hnsw index that returns the right shape, a distance metric that rejects bad input, an integration test that goes green — so you can stop at any point and still have a vector similarity search system that works for the cases covered so far.

What you will build

The finished vector similarity search system exposes five capabilities, each with a focused test:

• HNSW index
• Distance metric
• Insert pipeline
• Search reranking
• Metadata filter

HNSW index is the foundation that every later stage assumes; Distance metric is where vector similarity search correctness lives; Insert pipeline handles a high insert rate that degrades the graph before a background compaction can re-link nodes; Search reranking makes the system observable; and Metadata filter is the graceful-degradation layer that keeps overload from becoming corruption.

The build keeps the vector similarity search logic in ordinary modules you can read in one sitting, and pulls in dependencies only for the parts that are genuinely tedious to hand-roll. That keeps the diff between the tutorial and a real codebase small, and keeps the failure modes visible in code rather than hidden behind a framework.

Architecture and data model

The system is organized around HNSW index, Distance metric, and Insert pipeline, with Search reranking tying them together. Data flows as a flat array of fixed-dimension float vectors plus a graph overlay of neighbor links per node.

Each component is a module with a pure interface where practical: the same input produces the same output, and side effects are pushed to the edges. That discipline is what makes the phases independently testable, because a test for Insert pipeline does not need to stand up HNSW index and Distance metric to exercise the interesting branch.

The data model is deliberately explicit. A flat array of fixed-dimension float vectors plus a graph overlay of neighbor links per node Keeping the shape written down — in a schema, a type, or a fixture — means a change to the shape is a visible diff rather than a silent runtime shift, and it means the integration test can assert against a concrete shape instead of guessing.

Configuration and code are kept apart: the configuration says what vector similarity search should do, the modules say how. A reviewer can change a threshold or a policy without reading the implementation, which keeps the risky changes small and lets the diff stand as the review.

The failure path through Insert pipeline is designed before the happy path is finished, because a high insert rate that degrades the graph before a background compaction can re-link nodes is the case that distinguishes a vector similarity search system from a demo. Every component declares what it does when its inputs are wrong, and the integration test asserts each of those declarations.

Prerequisites and dependencies

Have these ready before you install anything, so the vector similarity search setup is not blocked by a missing tool:

• Node.js 20 LTS or later
• Familiarity with vector similarity search fundamentals
• A local environment to run the test suite

The runtime dependencies are chosen to keep the vector similarity search datapath legible:

• hnswlib-node (^3.0.0) — HNSW index construction and search
• fastify (^5.0.0) — Low-overhead query server
• pino (^9.0.0) — Structured logging

# setup for vector-embedding-store
pnpm init
pnpm add hnswlib-node fastify pino
pnpm add -D typescript tsx vitest

Versions are pinned to a major range so the vector similarity search install is reproducible; commit the lockfile. The hnsw index reads its operational parameters from the environment and validates them at startup, so a missing value fails the boot rather than the first vector similarity search request.

Directory structure

Keep each vector similarity search component in its own file so it can be tested without starting the others:

vector_embedding_store/
├── src/index.ts
├── src/hnsw-index.ts
├── src/distance-metric.ts
├── tests/vector_embedding_store.test.ts
└── package.json

Phase 1: HNSW index

Build the hnsw index that anchors vector similarity search and defines the core data shape.

// HNSW index: vector similarity search core
export function hnswindex(input: HnswindexInput): HnswindexResult {
  // Build the hnsw index that anchors vector similarity search and defines the core data shape.
  return processHnswindex(input);
}

After this phase the hnsw index works in isolation and is covered by a unit test. The hnsw index defines the data shape that every later stage assumes, so getting it right first means the later phases compile against a stable contract.

Treat the hnsw index as the single source of truth for the vector similarity search data shape. Validate inputs at the boundary and reject early, so a malformed value never reaches Distance metric or Insert pipeline where its effect would be harder to localize.

Phase 2: Distance metric

Add the distance metric so vector similarity search can finding the k nearest correctly.

// Distance metric — handles vector similarity search Add the distance metric so
export class Distancemetric {
  constructor(private opts: DistancemetricOptions) {}
  async run(ctx: Context): Promise<DistancemetricOutput> {
    return this.execute(ctx);
  }
}

After this phase the distance metric works in isolation and is covered by a unit test. This is where vector similarity search correctness lives: the distance metric is what makes finding the k nearest neighbors over millions of floating-point vectors without scanning every one true, and its test is the proof.

Keep the distance metric pure where you can. A function that takes an input and returns an output is trivial to test and impossible to mock wrong; a class with hidden state is the opposite. Push side effects to the caller and keep the distance metric about the transformation.

Phase 3: Insert pipeline

Wire the insert pipeline to handle a high insert rate without losing data.

// Insert pipeline: vector similarity search stage
export const insertpipeline = (next: Next) => async (req: Request) => {
  const out = await transform(req);
  return next(out);
};

After this phase the insert pipeline works in isolation and is covered by a unit test. The insert pipeline is the difference between a vector similarity search system that works on the happy path and one that survives a high insert rate that degrades the graph before a background compaction can re-link nodes.

Design the failure before the success. Decide what insert pipeline does when a high insert rate that degrades the graph before a background compaction can re-link nodes: reject, retry, shed, or degrade — and write that test first. A vector similarity search system whose failure behavior is unspecified will pick one at random under pressure, and it is usually the wrong one.

Phase 4: Search reranking

Integrate the search reranking to make vector similarity search observable and testable end to end.

// Search reranking for vector similarity search
export async function searchreranking(stream: AsyncIterable<Chunk>): Promise<Aggregate> {
  let acc = init();
  for await (const c of stream) acc = merge(acc, c);
  return acc;
}

After this phase the search reranking works in isolation and is covered by a unit test. Observability is not a finishing touch; it is how you know the vector similarity search system is doing what it claims in production.

Emit one structured event per meaningful action, with enough context to localize a failure without a second round-trip. The search reranking is what turns “something is slow” into “the distance metric is slow for inputs matching this shape,” which is the difference between an alert and an investigation.

Phase 5: Metadata filter

Finalize the metadata filter so the vector similarity search system degrades gracefully under overload.

// Metadata filter: vector similarity search finalization
export function metadatafilter(state: State): Result {
  validate(state);
  return finalize(state);
}

After this phase the metadata filter works in isolation and is covered by a unit test. The metadata filter is what keeps overload from becoming corruption: when the vector similarity search system is asked for more than it can do, it degrades in a defined way instead of failing in an undefined one.

Pick a degradation strategy for metadata filter — shed, queue, or backpressure — and make it explicit in code and in a test. Leaving the behavior to chance is how a vector similarity search traffic spike becomes a data loss incident.

Testing

Each phase has a focused test: hnsw index, distance metric, and insert pipeline are exercised in isolation, then an integration test drives the full vector similarity search pipeline and asserts the a high insert rate case is handled.

Start with unit tests that pin each component’s contract: the hnsw index returns the documented shape, the distance metric transforms correctly, and the insert pipeline fails in the documented way. These tests are fast and they localize regressions to the component that broke.

Then add an integration test that composes the components and drives a realistic vector similarity search scenario end to end. Assert the happy path first, then deliberately trigger a high insert rate that degrades the graph before a background compaction can re-link nodes and assert the system degrades rather than corrupts. Because the components are pure where practical, the integration test stays deterministic rather than flaky.

A vector similarity search suite must also prove the negative: bad input is rejected at the boundary, overload is shed rather than queued indefinitely, and a partial failure does not leave the system in an inconsistent state. These are the tests that catch the regressions a happy-path suite never would.

The entire vector similarity search suite runs with no external service; every dependency is either in-process or stubbed, so every case is reproducible in CI. That is what lets the insert pipeline failure-mode assertions live as tests rather than as a manual checklist that drifts out of date.

# run the vector similarity search test suite
pnpm vitest run

Deployment and operations

Run at least two vector similarity search instances behind a load balancer and share state in Redis. Set resource limits so vector similarity search overload surfaces as backpressure rather than failure.

Run at least two stateless instances behind a load balancer, and move any shared state into a shared store so the vector similarity search behavior is global rather than per instance. Size limits so the system sheds load before it runs out of memory, because a vector similarity search system that OOMs under load fails in a way that is hard to reason about.

Drain on deploy: stop accepting new work, let in-flight vector similarity search operations finish, then exit. Because the system holds no state that cannot be reconstructed, a rollout is a matter of cycling instances one at a time while watching the error rate.

When a dependency is unavailable, decide explicitly whether the vector similarity search system fails closed or degrades. Leaving that choice to chance is how a dependency outage becomes a silent data loss, so make the choice visible in code and in a test.

Security considerations

Validate all vector similarity search inputs at the boundary, reject oversized payloads before buffering them, and keep secrets in the environment rather than the repository.

The vector similarity search system is a choke point, so it is the right place to reject malformed input before it reaches distance metric or insert pipeline. Validate the shape and the size at the edge, and let the inner components assume what they receive is well-formed.

Never log secrets or full payloads; log enough to localize a failure without recording data that should not be on disk. Treat any external call as untrusted and validate its response before using it, because a vector similarity search system that trusts its inputs blindly is one malformed response away from corruption.

Keep dependencies pinned and run an audit in CI. A transitive dependency with a known vulnerability is still a vulnerability in your vector similarity search system, and the lockfile is what makes the audit meaningful.

Observability

Emit one structured event per vector similarity search operation with the inputs’ shape, the outcome, and a correlation id. Track error rate and latency per component and alert on a step change rather than a single spike.

The correlation id is the thread that ties a user-visible symptom to the vector similarity search decision and to the component log entry for the same operation. Without it, an investigation starts from a timestamp and a guess.

Watch the insert pipeline failure rate separately from the overall error rate. A jump in insert pipeline failures usually means a high insert rate that degrades the graph before a background compaction can re-link nodes, not a bug in distance metric, and the two have different owners and different fixes.

Chart the share of vector similarity search operations that shed, queue, or succeed. A change in the mix points at the cause faster than an aggregate latency can, because each class has a different root.

Troubleshooting

A high insert rate that degrades the graph before a background compaction can re-link nodes. Reproduce with a minimal fixture, then add a guard test so the vector similarity search system rejects the condition loudly.

Performance regresses under vector similarity search load. Profile with the sampling analyzer, locate the hot path in distance metric, and add a targeted benchmark to lock the regression.

Keep the vector similarity search dashboards separate from dependency dashboards. When an operation fails, the correlation id lets you decide in seconds whether the fault is hnsw index, distance metric, or insert pipeline, instead of guessing from a timeout.

Limitations and next steps

• Single-region; multi-region replication is out of scope.
• The hnsw index is in-process and not horizontally shared by default.
• No auth layer; assume deployment behind an edge that enforces identity.

The vector similarity search system is single-region by design; extending to multi-region means making hnsw index shareable across regions, which is a real change to the consistency model rather than a config flag. The insert pipeline is in-process, which is correct for one instance but must move to a shared store before scaling out.

There is no authentication or authorization layer; the system assumes it is deployed behind an edge that enforces identity. Adding one is straightforward but out of scope for this build, and doing it here would distract from the vector similarity search correctness that is the point.

The natural next step, once the vector similarity search pipeline is trusted in production, is to add a short-TTL cache for idempotent reads and a richer search reranking that exports histograms rather than counters. Both are additive and neither changes the contracts this build establishes.