Shaped vs Pinecone

Pinecone gives your agent 200 noisy chunks.
Shaped gives it the 10 that matter.

Replace your entire RAG and feature stack with one real-time retrieval engine. 100x fewer LLM tokens. Zero retry loops.

Try 200k queries free → See pricing →

Start with 200,000 queries free ($100 credit). No credit card required.

PINECONE Agent query Bad context Hallucinate ↺ Retry LLM Token Cost: $1.50 SHAPED Agent query Right context Agent answers Feedback loop LLM Token Cost: $0.03

PINECONE

Agent query
↓
Bad context
↓
Hallucinate
↓
↺ Retry
↓
LLM Token Cost: $1.50

SHAPED

Agent query
↓
Right context
↓
Agent answers
↓
Feedback loop
↓
LLM Token Cost: $0.03
70%

of Pinecone deployments use raw OpenAI embeddings that have never been updated since day one

Your retrieval is exactly as good as it was the day you set it up. Shaped changes that automatically.

What changes when you switch

Capability Pinecone Shaped
Knows who's asking ✕ User + session context
Learns from interactions ✕ Every query
Eliminates retry loops ✕ Right context first time
Data sources ✕ One index 30+ native connectors
Cost per answer ✕ ~$1.50 $0.03
Latency ✕ <100ms <50ms
Time to production ✕ Weeks 7 days
Try 200k queries free → See pricing →

Start with 200,000 queries free ($100 credit). No credit card required.

One query replaces your entire stack

Pinecone = Similarity Only.  Shaped = Relevance (Vector + BM25 Keyword + ML Reranking).

Pinecone stack
retrieval.py
import openai
import pinecone
import cohere
import redis
from your_db import get_user_rules, log_retrieval
from your_cache import build_cache_key
from your_filters import apply_business_rules

# 1. Generate embedding — separate API call, separate cost
embedding = openai.embeddings.create(
  model="text-embedding-3-small",
  input=query
).data[0].embedding

# 2. Query Pinecone — similarity only, no user context
results = index.query(
  vector=embedding,
  top_k=200,
  include_metadata=True
)
docs = [r.metadata for r in results.matches]

# 3. Keyword search — separate Elasticsearch call
# because Pinecone doesn't support BM25
es_results = es.search(
  index="products",
  body={"query": {"match": {"text": query}}}
)

# 4. Merge and deduplicate results manually
merged = merge_and_dedupe(docs, es_results)

# 5. Rerank with Cohere — another API call
reranked = co.rerank(
  query=query,
  documents=merged,
  model="rerank-english-v3.0",
  top_n=50
)
...
6 services ~400 LOC $1.50/query
Shaped
query.py
from shaped import RankQueryBuilder, TextSearch

query = (
  RankQueryBuilder()
  .from_entity('item')
  
  # 1. Retrieve — hybrid text search
  .retrieve([
    TextSearch(
      input_text_query='$query',
      mode={'type': 'vector', 'text_embedding_ref': 'your_text_embedding_name'},
      limit=100,
      name='vector_search'
    ),
    TextSearch(
      input_text_query='$query',
      mode={'type': 'lexical'},
      limit=100,
      name='bm25_search'
    )
  ])
  
  # 2. Filter — use WHERE clause
  .where('available = true AND region = $user_region')
  
  # 3. Score — trained model
  .score(
    value_model='click_through_rate_model',
    input_user_id='$user_id',
    input_interactions_item_ids='$interaction_item_ids'
  )
  
  # 4. Limit final results
  .limit(10)
  .build()
)

# 1 service. 1 query. Learns automatically. $0.03/answer.
1 API ~40 LOC $0.03/query
Try 200k queries free →

Start with 200,000 queries free ($100 credit). No credit card required.

Migrate in days, not months

Day 1

Connect your data sources

Day 2

First ShapedQL query running

Day 3

Results visibly better

Day 7

In production

Day 30

Results keep improving

RECOMMENDED

Start here

Connect your data, write one ShapedQL query, see results in a few hours.

Try 200k queries free →

Start with 200,000 queries free ($100 credit). No credit card required.

Proof of Concept is on us.

We give you $100 in free credits—enough for ~200,000 queries—so you can run a parallel test against Pinecone in your staging environment for zero cost.

Full replacement

Replace Pinecone, Cohere, Redis, and your ranking pipeline with a single ShapedQL query. Most teams are in production within 7 days.

Talk to us →

For your role

Engineers

One query replaces five services. Ships in 7 days. No infrastructure to maintain.

ML Engineers

Real models trained on your interaction data. Continuous learning. No retraining pipelines.

Data Scientists

Cold start handling built in. Full evaluation toolkit. Methodology documented.

Product Managers

Measurable uplift in engagement and revenue. Migration risk near zero. $100 to start.

+79% Retention
+9% Purchases
+10% Watch Time
+8% CTR

From teams in production on Shaped

Vox Outdoorsy Atmosfy SidelineSwap Sune Afterhour

Frequently asked questions

Yes. Retrieval, ranking, reranking, and the feedback loop. No separate embedding API call, no reranking service, no caching layer. One query, one cost.

Most teams have a working POC within a few hours and are fully in production within 7 days. Your $100 free credit covers the entire POC — no card required.

Most teams run their first query within a few hours and are in production within 7 days. Your $100 free credit covers the entire POC — no card required.

You'll know within a few hours. Roll back at any point — no lock-in, no migration required to get started.

Better results. Fewer tokens. Zero hallucinations.

$100 free credit · No card required · Live in 7 days

Start building free → See pricing →
SOC 2 Type II 99.95% uptime <50ms latency GDPR compliant