RELATIONAL TRANSFORMERS · A NEW MODEL ARCHITECTURE
The database
is the context.
Cells become tokens. Relationships carry attention. ASTERISM runs one
pretrained model over the whole graph — then predicts any target, zero-shot.
- DEPTH 01DATA
- DEPTH 02RELATIONS
- DEPTH 03CONTEXT
- DEPTH 04PREDICTION
THE PROMISE
Name the target.
Score the population.
One sentence names the missing cell. One pretrained relational
transformer fills it for every row. The task changes. The model
does not.
Which customers are likely to churn in the next 90 days?
AUROC 0.91 · re-scores as links landFlag fraudulent transactions as they arrive.
AUROC 0.97 · scored as they arriveWhat is each account worth over the next 12 months?
NMAE 0.24 · per-account estimatesHow many units will each SKU sell in the next 30 days?
NMAE 0.19 · every SKU, every storeWhich customers will buy again this month?
- 01cust_88410.92
- 02cust_12760.87
- 03cust_04130.81
Which devices will fail, and when?
NMAE 0.27 · days-to-failure, per deviceThe task changes. The model does not.
A yes or no. A number. A label. If the answer belongs to a row with history, declare the target and score the population.
THE PAYOFF · A PREDICTION YOU CAN INTERROGATE
The graph is the evidence. Remove a source, a relation, or a neighborhood. Re-score. Watch the answer move.
Carve one source, link, or subgraph out of the context and score again. If the metric falls, you found evidence the model was using.
Let ASTERISM remove each relation systematically. Rank every source and edge type by lift; tune breadth, depth, and recency on validation.
Slice the scores by lifecycle, cohort, source, or time. The model gives you a prediction. The population tells you what it means.
ONE IDEA TO CARRY AWAY
Attention can follow a join.
A transformer reads a sequence. A relational transformer reads a neighborhood.
Ninety seconds from now, that distinction will feel inevitable.
“Cells are tokens. Neighborhoods are context.
Attention is the join.”
We want to know whether customer #48291 will churn. Her orders live in one table, her refund in another, her support ticket in a third. A relational transformer does not flatten that world. It gathers it.
Her history enters the context with nearby customers whose outcomes are known. The churn cell is masked. Attention follows the relations. The model writes #48291 → 0.87.
No feature table was assembled. No model was trained for churn. The database supplied the context.
Cost follows context, not graph size.Every row receives a bounded neighborhood of a few thousand cells. A graph with 46 billion edges costs the same per row as one with 46 thousand. That is the route to scale.
It has never seen your database.The transformer was pretrained across hundreds of other databases. Yours arrives as context. It reads the schema cold, follows the relations, and predicts zero-shot.
FOUR MOVES. THAT IS THE METHOD.
Rows become nodes. Foreign keys become edges. Every cell becomes a token that knows its value, column, table, and place in the graph. Nothing is flattened first.
For each row, the sampler gathers its history and labeled peers into one bounded context. Every fact is sampled as of that row's timestamp, so the future cannot leak backward.
Cells attend within rows, across links, back through references, and down columns. The join, the aggregate, and the comparison become learned attention paths.
Which customers will churn in the next 90 days?The pretrained model fills the masked target for the whole population. New links arrive. Context changes. The rows score again.
WHAT A RELATIONAL TRANSFORMER IS — AND IS NOT
No giant join turns the database into one frozen matrix. The model keeps the rows, columns, and relationships visible all the way through attention.
Churn does not get one architecture and fraud another. The target changes; the pretrained relational transformer and its learned method stay the same.
Values are tokens. Rows and columns provide position. Relations determine which evidence can meet. The database becomes the model's context.
NOW MAKE THE METHOD LIVE
Asterism turns relational transformers
into a database that keeps learning.
- 01CONNECT
- 02DISCOVER
- 03COMPOUND
- 04PREDICT
Bring the world
as it is
Point ASTERISM at the sources
you already run.
Rows become nodes.
Foreign keys become edges.
The schema becomes structure.
No feature table first.
No graph to hand-build.
SQL / DATABASES
WAREHOUSES / LAKES
EVENT STREAMS / LOGS
APIS / WEBHOOKS
DOCS / EMAIL
CHAT / TRANSCRIPTS
OBJECT STORAGE
SAAS / CRM
150+ CONNECTORS
CONNECTOR LAYER
Find the links
you never wrote
The declared schema is only
the beginning.
ASTERISM discovers entities,
foreign keys, joins, and links
hidden across sources.
Every recovered relation
becomes a new attention path.
orders_container_id ↔
customers_id
confidence: 0.92
type: foreign_key
product_mentions ↔
support_tickets
confidence: 0.88
first_seen: 2024-04-21
Ticket
confidence: 0.96
source: chat transcripts
conversations.user_id↔
billing.customer_id
confidence: 0.94
Teach it once.
Rewrite the past.
Add a parser today.
Run it over every document,
message, and event already stored.
The extracted facts return
as first-class relations.
Old history gains new links.
Every future context gets richer.
PARSERS ADDED OVER TIME
HISTORYEXTRACTED FROM CONVERSATIONS
The Snowflake booth was packed.
We had 200+ demos.
Did the Snowflake booth raise sales?
snowflake_booth → sales_influenced
confidence: 0.91
evidence: conversation_27491
date: 2024-04-21
Ask once.
Score everyone.
Name the target in one sentence.
The pretrained transformer
masks that cell for every row
and fills the population.
Contexts come from cache.
Compute follows rows scored,
not the size of the graph.
PREDICTION RUN
Which customers are likely to churn in the next 90 days?
● Model: rt-j · pretrained · zero-shot
Fine-tuning: none
● Context: BFS over links · 8,192 cells/row
walk-ranked neighbors · latest first
● Context tuning: grid on validation
best (breadth, depth) · ensemble ×4
● Output: 2.4M customers scored
AUC 0.91 calibration 0.98
● Serving
– 1.2M predictions/sec · p95 38ms
– embeddings served from cache
– re-scores as new links land
CHURN RISK
TOP 10 CUSTOMERS LIKELY TO CHURN
RISK SCORE