Shaped | The Context Engine for Agents

What Shaped looks like in production.

Your agent asks Shaped for context. Here's what comes back.

WITH SHAPED

"How do I configure SSO for enterprise accounts?"

"How do I add middleware to the Express app?"

"Find me running shoes for trail running under $150"

ShapedQL

SELECT content, title
FROM
  text_search('SSO enterprise configuration', mode=vector),
  text_search('SSO enterprise', mode=lexical),
  similarity(user_id=$user_id)
WHERE doc_type = 'guide'
ORDER BY relevance(user, item)
LIMIT 10

ShapedQL

SELECT file_path, content
FROM
  text_search('middleware express', mode=vector),
  text_search('middleware express', mode=lexical),
  file_recency(repo_id='acme-api')
WHERE language = 'typescript'
ORDER BY relevance(user, item)
LIMIT 10

ShapedQL

SELECT product_id, name, price, description
FROM
  text_search('trail running shoes', mode=vector),
  similarity(user_id=$user_id)
WHERE category = 'running-shoes'
  AND price <= 150
ORDER BY relevance(user, item)
LIMIT 10

1.docs/enterprise/sso-setup.md0.97

2.docs/enterprise/saml-configuration.md0.94

3.docs/auth/identity-providers.md0.91

4.docs/admin/org-settings.md0.87

5.changelog/v2.4-sso-updates.md0.83

1.src/middleware/auth.ts0.96

2.src/middleware/logging.ts0.93

3.src/app.ts (lines 24-41)0.89

4.docs/middleware-guide.md0.85

5.src/routes/api.ts (lines 8-15)0.82

1.Nike Pegasus Trail 5 - $1290.95

2.Hoka Speedgoat 6 - $1450.93

3.Salomon Sense Ride 5 - $1390.90

4.Brooks Catamount 3 - $1490.87

5.New Balance Fresh Foam Hierro - $1340.84

Also available via Python SDK, TypeScript SDK, or MCP. See docs →

WHAT THE AGENT RETURNS:

Based on the SSO setup guide and SAML config docs:

1. Navigate to Admin → Org Settings → Authentication
2. Select your identity provider (Okta, Azure AD, Google)
3. Upload your SAML metadata XML

See docs/enterprise/sso-setup.md for full walkthrough.

WHAT THE AGENT RETURNS:

// Following patterns from auth.ts and logging.ts
import { authMiddleware } from './middleware/auth';
import { requestLogger } from './middleware/logging';

app.use('/api', authMiddleware);
app.use('/api', requestLogger);
app.use('/api', yourNewMiddleware);

WHAT THE AGENT RETURNS:

Based on your running history and past purchases:

1. Hoka Speedgoat 6 ($145) — You loved the Speedgoat 5. Improved grip, same cushion.
2. Salomon Sense Ride 5 ($139) — Great for mixed terrain.
3. Nike Pegasus Trail 5 ($129) — Best value for road-to-trail.

WITHOUT SHAPED

"How do I configure SSO for enterprise accounts?"

"How do I add middleware to the Express app?"

"Find me running shoes for trail running under $150"

Python

import pinecone
import cohere
from openai import OpenAI

# 1. Embed user query
res = client.embeddings.create(
    input=query, model="text-embedding-3-small")
query_vec = res.data[0].embedding

# 2. Vector search (200 noisy docs)
idx = pinecone.Index("my-index")
raw_docs = idx.query(
    vector=query_vec, top_k=200, 
    include_metadata=True)

# 3. Rerank with static model
co = cohere.Client('API_KEY')
reranked = co.rerank(
    query=query, 
    documents=[d.metadata['text'] for d in raw_docs['matches']], 
    top_n=10)

# 4. Stuff massive context and hope
context = "\n".join(
    [doc.document['text'] for doc in reranked.results])
prompt = f"Context: {context}\n\nAnswer: {query}"

Python

import pinecone
from openai import OpenAI

# 1. Embed query
res = client.embeddings.create(input=query, model="text-embedding-3-small")
query_vec = res.data[0].embedding

# 2. Pure Vector Search 
# (Embeddings are notoriously bad at exact code syntax)
idx = pinecone.Index("codebase-index")
raw_chunks = idx.query(vector=query_vec, top_k=200, include_metadata=True)

# 3. No BM25 / Lexical fallback
# Misses exact file paths like 'src/middleware/auth.ts' 
# because semantic meaning != exact code syntax.

# 4. Blind chunking destroys context
# Stuff 200 random 512-token chunks of code into the prompt 
# and hope the LLM can piece the file back together.
context = "\n---\n".join([c.metadata['code_snippet'] for c in raw_chunks['matches']])
prompt = f"Codebase:\n{context}\n\nAnswer: {query}"

Python

import pinecone
from openai import OpenAI

# 1. Embed query 
# (Embeddings often lose strict intent like "under $150")
res = client.embeddings.create(input=query, model="text-embedding-3-small")
query_vec = res.data[0].embedding

# 2. Vector search with clumsy metadata filters
idx = pinecone.Index("product-catalog")
raw_docs = idx.query(
    vector=query_vec, 
    top_k=200, 
    filter={"price": {"$lte": 150}},
    include_metadata=True)

# 3. Personalization?
# Fetching past purchases from Redis to manually rerank 
# 200 items in Python takes too long.
# user_history = redis.get(user_id) # Skipping due to latency...

# 4. Stuff generic, unpersonalized products into prompt
context = "\n".join([doc.metadata['desc'] for doc in raw_docs['matches']])
prompt = f"Context: {context}\n\nRecommend shoes: {query}"

50,000 TOKENS OF CONTEXT

90% irrelevant. $0.50/query. Static forever.

WHAT THE LLM RETURNS:

SSO (Single Sign-On) allows users to authenticate using a single set of credentials. Enterprise accounts can configure SSO through the admin panel...

// Generic overview with no specific steps

WHAT THE LLM RETURNS:

// Generic Express.js middleware example
app.use((req, res, next) => {
  console.log('Request received');
  next();
});

// No reference to your actual codebase patterns...

WHAT THE LLM RETURNS:

Here are some popular running shoes:
- Nike Air Zoom Pegasus 41 ($130)
- Adidas Ultraboost Light ($190)
- Asics Gel-Kayano 31 ($160)

// Generic bestsellers, not trail-specific, not personalized, wrong prices...

10 results · 2,100 tokens · 38ms · Your agent gets exactly what it needs — nothing more.

This is how most agents retrieve context today.

Query

Embed

Vector DB

Reranker

Stuff into prompt

LLM

200 results by cosine similarity

Static model doesn't learn

50,000 tokens $0.50/query · 90% noise

Too much irrelevant context

200 results. 190 are noise. Your agent re-retrieves, burning tokens and time.

No personalization

Senior engineers and new hires get the same 200 chunks.

Never improves

Day 100 is no smarter than day 1.

This is how agents retrieve context with Shaped.

Query + User

Shaped

10 ranked results

LLM

Learns from every interaction

Stop manually tuning. Shaped ingests clicks and other signals
to tune your models in realtime.

Ranked, minimal context

Shaped returns only relevant results. 100% of the context matters.

No retry loops

Shaped gets it right the first time. Your agent doesn't re-retrieve.

Learns from every interaction

User rephrases or thumbs-down? Day 100 is dramatically better than day 1.

0x fewer tokens 0ms latency No retry loops Learns from feedback

One API. Not five services.

Vector DB (e.g. Pinecone)

Search engine (e.g. Elasticsearch)

Reranker (e.g. Cohere)

Feature store (e.g. Redis)

Glue code (~2,000 LOC)

5 services5 bills5 points of failure

Shaped 1 API · ~40 LOC

1 service1 bill1 endpoint

Capability	DIY Retrieval Stack	Shaped
Retrieval	Single embedding space	Multi-retriever (vector + lexical + behavioral)
Ranking	Static reranker model	ML models that learn from outcomes
Personalization	None	User-aware ranking via user_id
Context size	50K+ tokens (top-k dump)	2,500 tokens (ranked LIMIT 10)
Infrastructure	Vector DB + Search engine + Reranker + Feature store + glue	1 API call
Improves over time	No - static from day 1	Yes - retrains on agent feedback
Query language	Multiple SDKs + custom code	ShapedQL (SQL-like, declarative)