HNSW: the algorithm powering fast vector search
hierarchy = speed
• sparse layers (express lanes)
• dense bottom (local streets)
96% recall, 44x faster than brute force
shaped.ai/blog/hnsw-explaine…
We just shipped the Signal Engine.
Define a feature once. Get training, serving, and streaming for free.
25+ signal primitives
<5ms serving
Point-in-Time correctness by default
Change a feature, see evaluation results in hours
Blog post: shaped.ai/blog/introducing-t…
Signal primitives ship today: aggregations, ratios, cross features, vector similarity, geo distance, text similarity, cyclic time encoding, sequences for SASRec/BERT4Rec, and more.
It's not just for ranking models. Signals power search relevance, agent context enrichment, faceted filtering, and any workload that needs fresh features at query time.
Docs: docs.shaped.ai/docs/v2/engin…
We just shipped an MCP server.
pip install shaped-mcp
Works natively with Cursor, Claude Code, Windsurf, VS Code Copilot, Gemini, OpenAI — any MCP-compatible agent.
Add it to your config. Your agent has better retrieval instantly. No custom integration code.
$100 free credits. No credit card.
→ shaped.ai/agent-context
If you're building agents and retrieval quality is killing you, try it. We'd love your feedback.
Netflix doesn't just personalize which movies you see. It personalizes which ROWS you see.
We wrote the playbook on how to build it.
shaped.ai/blog/how-to-build-…
discover weekly = hybrid filtering
collaborative: fails for new releases
content-based: too obvious
hybrid: ELSA + AI enrichment + adaptive scoring
shaped.ai/blog/how-to-build-…
"the new model feels better"
that's not a metric
offline eval:
Recall@10: 0.45 → 0.50 (+11%)
NDCG@10: 0.54 → 0.61 (+13%)
online A/B test:
CTR: +15%
Conversion: +19%
now you have proof
shaped.ai/blog/ab-testing-re…
your agent's config:
• ranking formula in .env
• filters hardcoded
• vector DB settings in UI
• features scattered
when it breaks: 🤷
with GitOps:
• everything in Git
• PR review before deploy
• one command to rollback
• clear audit trail
shaped.ai/blog/gitops-for-ai…
your agent: 5 seconds to respond
your LLM: 3 seconds
where are the other 2 seconds?
retrieval.
vector DB: 220ms
filtering: 50ms
scoring: 120ms
reordering: 60ms
network hops: 4x
= 450ms+ per query
Shaped's fast_tier: 30-100ms for all 4 stages unified
shaped.ai/blog/sub-100ms-dis…
The scaling law for agents isn't model size. It's what goes into the context window.
Attention is quadratic. 2x tokens = 4x cost.
10 ranked results > 200 stuffed chunks. Every time. At 10x lower cost.
We wrote the math:
shaped.ai/blog/context-windo…
your agent: "order has shipping tier 4 and status 7"
what the customer needed: "ships in 2 days via FedEx, in transit"
AI Views fix this—enrich at write time, not read time
shaped.ai/blog/the-status-co…
most "AI hallucinations" are just stale retrieval
user: "is this in stock?"
agent's index: last updated 2am
product: sold out at 9am
agent: "yes"
the model is fine. your data pipeline is broken.
how to fix it 👇
shaped.ai/blog/why-your-ai-a…
Your agent just cost you $40,000.
It recommended wireless headphones to a customer who bought the exact same pair yesterday.
The problem? It optimized for semantic similarity. It ignored business logic.
Here's how to fix it 🧵
Real production benchmarks:
Traditional stack:
→ 250-500ms (3-4 network hops)
→ Hours to update weights
→ 42% precision@5
Value models:
→ 50-100ms (single query)
→ Seconds to update weights
→ 68% precision@5
Same data. Better architecture.
We built a complete technical guide:
✓ Working code
✓ Architecture diagrams
✓ Dynamic weight tuning patterns
✓ Real-world use cases (e-commerce, travel, content) ✓ When NOT to use this approach
Read it here: shaped.ai/blog/building-mult…
