Agentic AI Capability

One question can read three systems at once.

A question about open purchase order lines reaches three systems: the orders in SAP, the orders of an acquired division in Oracle, and the weekly plan in a planning suite.

THE QUESTIONWhich open PO lines can I still cancel this week, and whatare they worth?LTWHAT HAPPENS UNDERNEATH01Read the questionthe business objects andthe period are recognised02Find the tablesthe ontology gives themapped table in each system03Split by sourceone query per system, inthat system’s own language04Read in placeeach system runs its partand returns its own rows05Join and presentthe engine matches onshared keys and appliesaccessREAD IN PLACESAP S/4HANAOracle FusionPlanningsuiteTHE ANSWER38 lines · $412,00026 lines from SAP, 12 from Oraclecover checked against the weekly plan4.2 s · three systems read in placeshowcase estate · the three systems are read at the same time
The question at the top, the five steps underneath, and the three systems that answered. The rows stay in the systems that hold them. Showcase estate.
4.2 s
average time to an answer across three systems in the showcase estate
showcase estate
3
systems read for one answer, joined in the federated query engine
showcase estate
In place
rows are read in the system that holds them and returned to the asker
product
2
sets of access rights apply to every row: the connection and the person
product
Where this sits

Stage 1 of six: register

Step by step

What happens between the question and the answer

Two things travel: the query out to each system, and the result back.

THE QUESTIONPLANSPLITJOINPRESENTThe person orthe agentLima federatedquery engineYour sourcesystemsA question or arulefrom a person or anagentPlan the worktables, filters,join keysOne query persourcein each system’slanguageEach source runsits partwith its own indexesJoin and filtermatched on sharedkeysThe answer andthe querya table, a chart, asummaryaskplansendrowspresentthe rows stay in the system that holds them · the plan and the description live in Lima
The middle lane is the federated query engine. It plans the work, writes one query per system, and matches what comes back on the keys your mapping confirmed.

A question arrives from ChatLima, from an agent run or from a report. Juno recognises the business objects in it: purchase order lines, a plant, a week, a value. The mapping made at the Map stage says which table in which system carries each one.

From that, Juno writes a query plan: the tables to read, the filter for each source, the keys that join the parts and the order of the work. The plan is kept with the answer.

Each source then receives a query in its own language, holding the narrowest filter Juno can state: one plant, one week, open lines only. The source runs it with its own indexes and returns a small result.

The federated query engine matches the parts on the shared keys, applies the row and column rules of the person asking, and presents the result. Every row stayed in its own system; the query and the result travelled.

On screen

The query plan behind one answer

One line per step, with the source, the filter, the result and the time.

Query plan · “open PO lines I can still cancel this week” · showcase estate
StepWhere it runsWhat it readsWhat comes backTime
1 · Read and planLimathe ontology, the mapped tables and the filtersa query plan1.9 s
2 · Open linesSAP S/4HANAEKKO · EKPO · EKET, plant 1210, open lines26 rows1.8 s
3 · Open linesOracle Fusionpurchase order lines of the acquired division12 rows1.4 s
4 · CoverPlanning suiteplanned orders and demand for 38 materials38 rows0.9 s
5 · JoinLima federated query enginematched on material, plant and supplier38 rows0.3 s
6 · Access and presentLimarow and column rules for L. Tan, supply planner38 rows0.2 s

Steps 2, 3 and 4 run at the same time, so the longest of them sets the reading time. Showcase estate.

The plan is the audit record of an answer. It names each source, the filter sent to it, the rows it returned and the time it took. When an auditor asks where a number came from, this is what answers.

Speed

Four habits keep the answer quick

Response time comes from how much each source is asked to do, and from how much of the work runs together.

WHERE THE 4.2 SECONDS WENTRead the question and plan the work1.9 sRead the three systems, at the same time1.8 sJoin the parts on the shared keys0.3 sApply access and present the answer0.2 sshowcase estate · the longest source read sets the reading time
Reading the three systems took 1.8 seconds because they were read at the same time; the longest single read sets that figure. Illustrative values; the mechanics are the product’s.
The filter goes to the sourceA plant, a week and an open status travel with the query, so a system holding millions of rows returns tens.
Only mapped columns are readThe query asks for the columns the question needs, so wide tables return narrow results.
The sources are read at the same timeThree reads run together, so the longest of them sets the reading time.
A follow-up reuses the planA follow-up that narrows the same result keeps the plan and the context, so the second answer arrives faster.
A governed copy

When reading a copy in your lake serves you better

Reading in place is the default; a copy earns its place for history, for slow sources and for heavy repeated reads.

Years of history in one questionSources keep a limited window online. The Lima Quanta lands the older periods in your lake, and Juno reads them there.
A source that answers slowlySeveral planning suites publish an extract on a schedule. A copy of it gives every reader a steady response time.
A heavy read on a busy systemMonth-end is a busy week for an ERP. A copy carries the repeated heavy reads.
The same owner and the same rulesA governed copy keeps the owner, the access rule, the refresh and the version history of its data set.

The read mode belongs to the data set, and one project can hold all three: read in place, a governed copy in your lake, and a file drop or a stream. Your team sets the mode when the data set is registered and can change it later.

Reading in place gives the state of the source at the moment of the question. A governed copy gives a fixed refresh, a steady response time and years of history in one place. Either way the ontology, the access rules and the audit record stay the same.

The running example

Wednesday 09:12: one question, three systems, 4.2 seconds

Showcase estate.

The running example

L. Tan asked ChatLima which open purchase order lines she could still cancel this week, and what they were worth. The purchasing data for her plant sits in SAP; the acquired division’s orders sit in Oracle Fusion; the cover for each material comes from the planning suite.

Juno read the question, found the mapped tables for each system, and sent one query to each with the plant, the week and the open status in the filter. SAP returned 26 lines, Oracle 12, the planning suite the cover for 38 materials. The engine joined them on material, plant and supplier, applied L. Tan’s access rights and presented 38 lines worth $412,000 in 4.2 s, with the query on show.

Next capability

Keeping up when a source changes

What a sync scan finds when a column is added, a table is renamed or a field goes away, and what happens to the mappings and rules that depend on it.

How this is usually done

A copy of everything, before the first question

The effort moves from moving data to describing it.

The common route to a cross-system answer is a central copy. Every source gets a pipeline, the tables land in a warehouse, a modelling layer gives them meaning, and the question is answered from the copy. Each new source adds a pipeline to keep current, and the answer is as fresh as the last load.

Juno asks the systems directly and keeps the description, the mapping and the plan. A new source becomes answerable once it is registered and mapped. Where a copy serves the work better, Quanta lands a governed one under the same rules.

Terms on this page
Federated query engineThe part of Lima that plans one question across several systems and joins what each returns.
Read in placeQuerying rows in the system that holds them, so the rows stay in the source.
Query planThe steps Juno works out for one question: the tables, the filters, the join keys and the order.
Filter at the sourceSending the narrowest condition with the query, so the source returns a small result.
Join keyA value two systems share, such as a material number, used to match their rows.
Governed copyA scheduled copy of a source in your own lake, with its owner, its refresh and its access rule.
Questions
How many systems can one question read?

Any registered data set in the project is available to a question. In the showcase estate three systems answer a typical planning question; production estates connect more than 40 source systems.

What happens when one source is busy?

Juno returns the parts it has, names the source it is waiting for and shows the time of its last good read. The person decides to wait for the full answer or to work with the parts that arrived.

Do our rows end up inside Lima?

Lima keeps the description of each table, the query plan and the record of each run. Rows are read for the question at hand and go to the person who asked. A governed copy lands in your own lake when your team chooses that mode.

Does this work for agents as well as for questions?

Yes. An agent run reads through the same engine, with the same mapping and the same access rules. The rules engine evaluates its conditions on the rows that come back and records the plan with the run.

In the platform

Where Juno sits among the four capabilities

Each one builds on the one before it.

One shared foundationCatalogLineageAccess policyWrite path

See Juno answer and act on your own data

Bring two systems and one decision your planners make every week. In the first week we register the data, confirm the mapping with your team, answer your questions in ChatLima and run the first agent at the level you choose.