Workshop 8: Enrichment Rules

Some facts the business relies on are written down nowhere. No SOW says “this was a healthcare project”, but everyone knows it was, because the client is a hospital system. No resume lists every skill a project used, but the project description makes it obvious. An enrichment rule writes these facts into the graph after ingestion. It reads a node (and, if you want, its neighbours), follows a plain-language instruction, and writes back an attribute, a node, a relationship, or both. This workshop decides which of those business rules are worth automating, and writes each one so it can be tested. Run it in Phase 3, after the first pilot ingestion. Rules written against an empty graph are guesses. With real nodes on screen, the room can see which gaps actually hurt the golden questions.

At a glance


Purpose	Identify implicit business rules; decide which become enrichment rules; write, test and schedule them
When	Phase 3 (Pilot ingest), weeks 5–6, after structured data and a document sample are loaded
Duration	90 minutes, plus a 30-minute follow-up to review test results
Attendees	Experio: FFE (leads), Experio SMEs as needed. Client: PM, super user, the SME for each use case
Inputs	Pilot graph with real data; golden questions and their current results; taxonomies loaded
Outputs	Rule worksheet (one row per rule); rules saved and previewed; a Flow that runs them after ingestion
Admin pages	Model & Define > Enrichment Rules, Process > Flows, Process > Flow Executions

Before the workshop

FFE and Experio SMEs

Run the golden questions against the pilot graph. For each partial or failed answer, note whether the missing fact is in a source but not extracted (fix extraction) or not stated anywhere (enrichment candidate).
Pick five real nodes per candidate target type to test on during the session.
Check the taxonomies the rules will use are loaded and their leaf values are active.

Client homework (SMEs)

Bring the “everybody knows” rules for your use case: how you’d tag, group or connect things if you did it by hand in a spreadsheet.
For each rule, bring two or three examples with the answer you’d expect.

Which tool for which fact

Ask this for every candidate. Most “rules” turn out to belong somewhere else.

If the fact is…	Use	Example
Written in the document text	Extraction instructions on the ontology attribute or artifact type	Contract `expiration_date` from the SOW
Held in a system of record	Data mapping (Workshop 6)	Client `industry` from Salesforce
Not stated anywhere, but derivable from what’s in the graph	Enrichment rule	Project industry from its client’s industry
About how to read or phrase an answer at question time	Assistant instruction (AI Instructions, Cypher instructions)	“Engagement lead means WORKS_ON_PROJECT role ‘Engagement Lead‘“
Relative to today (due soon, expiring, recent)	Leave it to the question; Cypher compares dates when asked	”Which SOWs expire in the next 90 days?”

Prefer the source over a rule. If a system of record holds the fact, map it. A rule is an informed guess by a model; a mapped column is the firm’s own record.

Agenda

Time	Item
0:00	Recap pilot results: which golden questions failed and why (15 min)
0:15	Collect candidate rules from SMEs (15 min)
0:30	Sort each candidate with the decision table (15 min)
0:45	Write the surviving rules in the worksheet (25 min)
1:10	Preview one rule live on real nodes (15 min)
1:25	Agree how and when rules run (5 min)

Running the workshop

For each rule, fill in one row of the worksheet:

Field	Question to ask the room
Target entity type	”Which kind of node gets the new fact?”
Target filter	”All of them, or only some?” For example, only nodes where the output attribute is null
Input mode	”What would you look at to decide?” Only the node’s own attributes (Selected Attributes), the whole node (Full Node), or the node plus everything one hop away (Neighborhood)
Instruction	”Say the rule as you’d explain it to a new analyst.” Write it in plain language
Controlled output	”Is the answer a free choice or one of a list?” If a list, use `@TaxonomyName`
Output	”What should change in the graph?” An attribute, a new node, a relationship to an existing node, or a new node and relationship
Expected results	”For these five nodes, what’s the right answer?”

Notes for writing the prompt:

Placeholders. In Selected Attributes mode, use {attribute_name} (for example {description}). In Full Node and Neighborhood mode, {node} inserts the whole node. In Neighborhood mode, {related.Label.attr} inserts an attribute from each related node of that type, such as {related.Client.industry}.
Taxonomies. @Industry expands to the active leaf values of the Industry taxonomy, so the model chooses from your list rather than inventing labels.
Relationship outputs link to existing nodes only. For a Relationship output, Experio looks up the target node by a match attribute (default name), either Exact match only or Exact, then vector similarity (the second needs a vector index on the target type). If no node matches, that link is skipped. Use Node + relationship when the target should be created.
No “today”. The rule prompt is not given the current date. Rules that depend on today’s date go stale the day after they run. Leave those to the question.

Testing

Open the rule’s Preview Rule page and run it on a small sample of nodes. The preview makes no changes to the graph. Compare the output with the expected results on the worksheet. Fix the prompt and preview again until the sample is right. Then run the job, and spot-check results through lineage on the node.

Running

Enrichment rules do not run automatically per file. They run when you start them:

On demand: Run Job from Model & Define > Enrichment Rules. Turn on Overwrite existing values only when you want to re-process nodes that already have a result.
After ingestion: add an Enrichment step to a Flow in Process > Flows, after the scan and ingestion steps, and schedule it. This is the normal setup after go-live, so new documents are enriched on each refresh.

Every write is recorded in lineage, so users can see which rule and model produced a value.

Worked example: Northbridge Consulting

The pilot showed golden questions Q1 (“cloud migration projects for healthcare clients”) and Q4 (“Azure data platform experience in healthcare”) returning partial answers. Projects had no industry, and skills lived only on people. Three rules came out of the session.

Rule 1: Tag project industry

Field	Value
Target	Project; filter `industry` is null
Input mode	Neighborhood
Output	Attribute `industry` (the team first added this attribute to Project in the ontology)

You are tagging a consulting project with the industry it served.

Project:
{node}

Industry of the client this project was for:
{related.Client.industry}

Choose exactly one industry from this list: @Industry
Use the client's industry when it is present. If it is missing, decide from
the project name and description. If you cannot tell, return null.

Rule 2: Infer skills used

Field	Value
Target	Project; filter `description` is not null
Input mode	Selected Attributes (`name`, `description`)
Output	Relationship Project `USED_SKILL` Skill, match on `name`, Exact match only

The Project description is filled from the SOW scope during extraction, so this rule reads the SOW’s scope indirectly.

Project name: {name}
Project description and scope: {description}

List the skills this project clearly required. Choose only from: @Skills
Return at most 8 skills. Do not include a skill unless the description gives
direct evidence for it.

Because this is a Relationship output, it links only to Skill nodes that already exist with exactly that name. Skill nodes come from resumes, which are extracted with @Skills in their instructions, so names line up. Skills no one at Northbridge lists on a resume are skipped. The room accepted that.

Rule 3: Obligation due soon flag (not adopted)

Helen Park asked for due_within_90_days on Obligation. The team wrote it out:

Obligation: {description}
Due date: {due_date}
Is this obligation due within 90 days of today? Answer true or false.

It was dropped. The prompt has no reliable “today”, and the flag would be wrong a day after each run. Instead, golden question Q7 is answered at question time, where Cypher compares due_date with the current date. The rule stayed on the worksheet with the reason, so nobody proposes it again.

Running at Northbridge

Marcus built a Flow “Nightly refresh”: scan SharePoint → ingestion → Enrichment (Rule 1) → Enrichment (Rule 2), scheduled nightly. Both rules keep their “is null” / “is not null” filters so re-runs process only new nodes.

Accelerate with AI

No AI helper drafts enrichment rules today. Admin Copilot (⌘J) can explain input modes, placeholders and output types from these docs while you write a rule. The best accelerator is the Preview Rule page. Iterate on real nodes rather than debating the prompt in the abstract.

Entering it in Experio

Create the rule

Model & Define > Enrichment Rules > Create Rule. Enter a name and description, choose the target node label, and add a target filter (attribute, operator, value; combine with AND or OR).

Set input, prompt and output

Choose the input mode, write the prompt with placeholders and @TaxonomyName tags, then configure the output type. Add per-attribute output instructions if the format matters (“title case”, “most specific match”).

Preview

From the rules list, open Preview Rule, set a small limit and run it. Nothing is written.

Run and schedule

Run Job once, and check results on the rule’s Jobs tab and in lineage. Then add an Enrichment step to a Flow in Process > Flows and schedule it.

See Enrichment Rules and Flows. Rules show a compatibility status. An invalid rule (for example, its target label or output attribute was removed from the ontology) cannot run until it is fixed and re-validated.

Common pitfalls

Writing a rule for a fact a system already holds. Map it instead.
Writing a rule for a fact the document states. Fix the extraction instructions instead.
Free-text outputs where a list exists. Without @TaxonomyName, the model invents near-duplicate labels (“Health Care”, “Healthcare Providers”).
Expecting a Relationship output to create targets. It links to existing nodes only. Use Node + relationship when targets must be created.
Rules that depend on today’s date. They go stale. Answer them at question time.
No target filter. Every run re-processes every node, costs more, and can overwrite reviewed values if Overwrite is on.
Assuming rules run per file. They run only on demand or in a Flow. Without a scheduled Flow, new documents are never enriched.

Exit criteria

Every candidate rule is sorted with the decision table, and the reason is recorded.
Each adopted rule has a worksheet row: target, filter, input mode, prompt, taxonomy, output, expected results.
Each rule is previewed on at least five real nodes and matches the expected results.
Each rule has run once, and results are spot-checked in lineage.
A scheduled Flow runs the rules after ingestion, in the right order.
The affected golden questions are re-run and their results recorded.

If the firm needs to restrict who sees what, continue to Workshop 9: Access Control. Otherwise go to Workshop 10: Assistants & Agent Flows.

Start Here

Plan

Workshops

Validate & Launch

Reference

At a glance

Before the workshop

Which tool for which fact

Agenda

Running the workshop

Testing

Running

Worked example: Northbridge Consulting

Rule 1: Tag project industry

Rule 2: Infer skills used

Rule 3: Obligation due soon flag (not adopted)

Running at Northbridge

Accelerate with AI

Entering it in Experio

Common pitfalls

Exit criteria

Next

​At a glance

​Before the workshop

​Which tool for which fact

​Agenda

​Running the workshop

​Testing

​Running

​Worked example: Northbridge Consulting

​Rule 1: Tag project industry

​Rule 2: Infer skills used

​Rule 3: Obligation due soon flag (not adopted)

​Running at Northbridge

​Accelerate with AI

​Entering it in Experio

​Common pitfalls

​Exit criteria

​Next

At a glance

Before the workshop

Which tool for which fact

Agenda

Running the workshop

Testing

Running

Worked example: Northbridge Consulting

Rule 1: Tag project industry

Rule 2: Infer skills used

Rule 3: Obligation due soon flag (not adopted)

Running at Northbridge

Accelerate with AI

Entering it in Experio

Common pitfalls

Exit criteria

Next