mufassall

Data & storage / 5 / 15

Vector search

An online store's help centre has 2 million article chunks, and customers ask in their own words: "can I get my money back for a broken kettle". Keyword search misses the answer when it uses other words. We add embeddings and cosine similarity, see why exact kNN is too slow, trade a little recall for milliseconds with HNSW, shrink memory with int8, handle filters, combine BM25 and vectors with reciprocal rank fusion, and feed the result to an LLM.

  • 14 min
  • Advanced
  • 11 steps
Recall@10 (golden set)
41 %
Search p95
12 ms
Index memory
—
Empty results
0 %
Open in Lab

  • request
  • reply
  • redelivery
  • duplicate
  • lost
  • call: the reply returns on the same line

Numbers are a model scenario, not measurements

The real request

Search API → Keyword index

Request

q = "singan choynak uchun pulimni qaytarib olsam bo'ladimi"
BM25 top 312
  1  "Choynak tanlash: 5 maslahat"         7.9
  2  "Elektr choynakni tozalash"           7.1
  3  "Singan ekran: ustaxona manzillari"   5.4
-- the answer is "Mablag'ni qaytarish: 14 kun ichida": no shared word3
  1. BM25 — Keyword ranking: it can only score words the text actually contains.
  2. top 3 — What the chat shows. All three are about kettles, none about refunds.
  3. no shared word — "pulimni" and "mablag'ni" mean the same money; for an index they are unrelated tokens.

Build it yourselfCheat sheetCompare lessons

Course: 74 lessons
  1. 01 Foundations

    1. The life of a request
    2. Latency and throughput
    3. Napkin math
  2. 02 Scaling

    1. Scaling up vs scaling out
    2. Load balancing
    3. Stateless servers
    4. Caching
    5. Cache invalidation
    6. CDN and edge caching
  3. 03 Data & storage

    1. SQL and NoSQL
    2. Database indexes
    3. Geo search: who is nearby
    4. Full-text search
    5. Vector search
    6. Storage engines: B-tree and LSM
    7. Bloom filters
    8. Database replication
    9. Sharding
    10. Consistent hashing
    11. Unique IDs
    12. CAP: when the link breaks
    13. Consensus and leader election
    14. Transactions
    15. File uploads
  4. 04 Communication

    1. API styles
    2. Message queues
    3. Pub/sub and event logs
    4. Real-time updates
    5. Reliable callbacks
    6. Sending webhooks reliably
  5. 05 Reliability

    1. Redundancy and failover
    2. Rate limiting
    3. Timeouts and retries
    4. Observability
    5. SLOs and error budgets
    6. Backpressure
  6. 06 Architecture

    1. Microservices
    2. Service discovery
    3. Sagas and the outbox
    4. Change data capture (CDC)
    5. Transactional outbox
    6. Distributed locks
    7. Batch vs stream
    8. Multi-tenancy
    9. Event sourcing
  7. 07 Security

    1. Authentication
    2. Authorization
    3. Encryption and secrets
    4. Secrets management
  8. 08 Shipping

    1. Testing
    2. CI/CD: from commit to production
    3. Zero-downtime migrations
    4. Feature flags
  9. 09 Whole systems

    1. From requirements to design
    2. Design a URL shortener
    3. Design a news feed
    4. Design a chat
    5. Design a video streaming site
    6. Design search autocomplete
    7. Design a job scheduler
  10. 10 Business processes

    1. Payment gateway
    2. Money and rounding
    3. Double-entry ledger
    4. Payment reconciliation
    5. Subscription billing
    6. Electronic digital signature
    7. Document lifecycle
    8. Audit log
    9. Notification delivery
    10. Bulk import
    11. Refunds and chargebacks
    12. Gapless document numbering
  11. 11 Case studies

    1. Case: how Discord stores messages
    2. Case: Figma multiplayer editing