Systems and records
Context graph missing links: a completeness audit for one workflow
By SourceX Editorial · Updated
10 sections
Short answer
To find the missing links that stop an agent using business records, audit one workflow from its decisions. Write the chain of records each decision must traverse, then test every hop on a sample of real cases: recorded, resolving to the right record, valid at decision time, visible to the agent and traceable to a source. Report complete paths per case, not coverage per relationship.
Checklist
Context graph missing links
Key takeaways
- Agents fail between tables, so audit the chain of records each decision needs rather than each table on its own.
- A case is complete only when every hop passes, so path completeness is lower than any single relationship's coverage.
- Write each path as a testable shape or query with a pass rule set before the counts are taken.
- Diagnose each failed hop as unrecorded, text-only, lost in export, hidden by permissions or stale, and fix it by cause.
- Keep inferred links labelled with evidence and confidence, and report completeness with and without them.
How do you find the missing links that stop an agent using business records?
Start from the decisions in one workflow, not from the schema. Write down each decision the agent has to make, the chain of records it must walk to make it, and then test that chain on a sample of real cases: is each hop recorded, does it resolve to the right record, was it true at the decision time, can the agent's identity see it, and does it trace back to a source. The missing links are the hops that fail, counted per hop and per case.
Per-case counting is the point. A context graph can have high coverage on every individual relationship type and still leave most decisions unanswerable, because a decision needs every hop in its chain at once. A refund decision that needs four hops fails whenever any one of them is missing, so the share of complete paths is lower than the weakest single hop, sometimes much lower.
This page gives a graph completeness audit for one workflow, a way to diagnose why each link is missing, and rules for links that have to be inferred. Which relationships an enterprise context graph should keep in general, and how to measure linkage quality across a whole multi-system dataset, are covered by the pages linked at the end.
Why does an agent stall on records that each look complete?
Because records are usually validated one table at a time, and agents fail between tables. Every ticket has a requester, every order has a customer and every invoice has an amount, so each export passes its own checks. The agent's question is different: for this ticket, which order, which shipment, which invoice, which payment and which policy version applied that day. If the ticket names the order only inside a free-text comment, the chain breaks on the first hop and the remaining records, however clean, are unreachable.
The failure is rarely loud. An agent that cannot reach the shipment record does not say so; it answers from what it can reach, such as the customer's most recent order, which is often a different one. That is why a completeness audit has to check that each hop resolves to the correct record, not only that some record is linked.
Business systems express these hops in different ways, which is part of the difficulty. Zendesk tickets carry a problem_id that links an incident to its problem ticket, and follow-up tickets are created from a closed ticket through a via_followup_source_id parameter. HubSpot connects records through associations, which can carry labels such as a paired parent and child, and marks one associated company as primary. Email threads link through Message-ID, In-Reply-To and References headers defined in RFC 5322. An audit has to know which mechanism each hop relies on.
Which workflow, and which decisions, should the audit cover first?
Pick one workflow that ends in a recorded outcome, has five to eight hops, and touches at least two systems. Refunds and credits, warranty claims, change requests and collections disputes all qualify. A workflow with a recorded outcome lets you check later whether the agent's decision matched what happened.
Then list the decisions inside it as questions the agent must answer from records. For a refund workflow: is this customer entitled to a refund, for which order line, under which policy version, and has a refund already been issued. Each question becomes a path. Keep the list to the decisions a person in the business actually makes; questions nobody asks produce hops nobody maintains, and auditing them inflates the missing-link count without improving the agent.
Agree the case sample with the workflow owner before measuring. A random sample of closed cases from a recent period is the default; add a stratum of escalated or reopened cases, because those are where links are likely to be missing and where an agent's mistakes cost the most.
How do you write the required path so it can be tested automatically?
Write each path as a shape that a validator can check case by case. W3C's SHACL language is designed for this: a property shape names a path, sequence paths chain several properties so one constraint can cover a multi-hop route, sh:minCount sets the minimum number of values that must exist, and the validation report lists each failure with its focus node and result path. That gives you exactly the per-case, per-hop failure list the audit needs, and the shapes become a reusable test you can run on every refresh.
If the records are not in RDF, the same structure works as a set of SQL or graph queries, one per hop, each returning the cases where the hop fails. What matters is that every hop has an explicit pass rule written down before the counts are taken.
Give provenance its own hop. W3C's PROV-O defines wasDerivedFrom for an entity built from a pre-existing one, wasAttributedTo for the agent an entity is ascribed to, and generatedAtTime for when an entity was completely created and available for use. A fact the agent relies on, such as a refund eligibility flag, should resolve to the record and the person or process that produced it; if it does not, the agent can use the fact but cannot justify the decision.
The graph completeness audit for a refund workflow
The sheet below is the audit for one illustrative workflow: a customer asks for a refund through the help desk, and the agent must decide eligibility and amount basis. Replace the systems and fields with your own and keep the columns. Each hop is tested on every sampled case; a case is complete only when every required hop passes.
| Hop | Link mechanism | Test on each case | Pass rule | Usual reason it fails |
|---|---|---|---|---|
| Ticket to customer | Requester and organization fields on the ticket | Requester resolves to one customer record in the CRM | Exactly one match on a stable ID | Requester used a personal email not on the account |
| Ticket to order | Order number in a custom field or linked record | Order ID present in a structured field and found in the order system | Found in a field, not only in comment text | Agent pasted the number into a comment |
| Order to shipment | Order ID on the shipment or delivery record | Shipment exists with a delivery status and date | One shipment per shipped line, dated | Split shipments recorded only on the first line |
| Order to invoice and payment | Invoice references the order; payment applied to invoice | Invoice and applied payment both resolve | Payment applied, not only received | Unapplied cash sitting on the customer account |
| Ticket to prior refunds | Earlier tickets, credit memos or problem links | Search finds any credit already issued for the same order line | Credit memos reference the order line | Credit issued in finance with no ticket reference |
| Decision to policy version | Policy document version in force on the request date | Policy version effective on the ticket's creation date is identifiable | Version and effective date recorded | Policy page overwritten with no history |
| Fact to source and author | Provenance of eligibility flags and notes | Each flag resolves to the record and person or job that set it | Source record and author present | Flag set by an import with no author |
| Hop visible to the agent | Permissions of the agent's integration user | Agent identity can read every record on the path | Same result through the agent's tools as through an admin query | Finance objects hidden from the support integration user |
Why is the link missing, and what is the right fix for each cause?
Classify every failed hop by cause before fixing anything, because the fixes differ and one of them is wrong more often than it looks: inferring the link. Five causes are listed below.
| Cause | How it shows up in the audit | Right fix | What not to do |
|---|---|---|---|
| Never recorded | No field, association or reference exists anywhere | Change the workflow so the link is captured at the moment of work | Backfill by guessing from names and dates and present it as recorded |
| Recorded only in text | Order number or ticket ID appears in a comment, email or note | Extract with a pattern, keep the source span, and mark the link as extracted | Discard the text after extraction, losing the evidence |
| Lost in export or de-identification | Link exists in the live system but not in the delivered data | Re-export with the join keys and apply one consistent pseudonym per identifier | Hash identifiers differently per table |
| Hidden by permissions | Admin query finds the hop; the agent's identity does not | Decide deliberately whether the agent should see it, and record the decision | Widen the integration user's access to make the audit pass |
| Stale or superseded | Link resolves, but to a record replaced before the decision time | Store validity time on the link and query as of the decision date | Overwrite the old link so history disappears |
Worked example: 80 refund cases from a help desk and an ERP (illustrative)
All figures here are invented for illustration. An evaluation team audits 80 closed refund tickets from one quarter, 60 drawn at random and 20 from reopened or escalated tickets. The audit runs the eight hops above against a sandbox copy of the help desk and the ERP.
Ticket to customer passes on 76 of 80 cases. Ticket to order passes on 52 in a structured field and on another 21 only through an order number in a comment, so the team extracts those with the source span kept and labels them extracted. Order to invoice and payment passes on 70. Ticket to prior refunds passes on 61, and the 19 failures cluster in the escalated stratum, where finance issued credits directly. Decision to policy version passes on all 80, because the policy page keeps dated versions.
With extracted links counted, 44 of 80 cases have a complete path; with recorded links only, 31 do. The team reports both numbers. The fix with the most effect is not in the graph at all: finance starts entering the ticket number on credit memos, which closes the hop that most often hid an earlier refund.
When should an inferred or remembered link be kept, flagged or dropped?
Keep an inferred link only when it carries its evidence and confidence, flag it whenever a decision depends on it, and drop it as soon as an authoritative record contradicts it. An agent cannot tell a recorded link from a guessed one unless the graph says so, and treating the two alike is how plausible but wrong context gets into decisions.
Time matters as much as confidence. Graphiti, an open-source framework for temporal context graphs, describes facts with validity windows and says that when information changes, old facts are invalidated rather than deleted, with every relationship traced back to the raw episodes that produced it. Apply the same rule to links in the audit: a link that was correct in March and superseded in May is not missing, but an agent deciding in June must not use it.
Some links should be dropped even when they are correct. If a record is deleted under a retention rule or a deletion request, links to it go too, and the audit should count the resulting gap as expected, not as a defect. Treat the rules for retention, erasure and access as inputs to the audit, set by the people responsible for them.
What the audit cannot tell you, and when to escalate
It measures whether the path exists and resolves correctly on a sample; it does not show that the agent will use the path well. Pair it with task-level evaluation that checks the agent's decisions against the recorded outcomes. It also inherits the sample's blind spots: a quarter of refunds may contain no chargebacks or partial returns, so the hops those cases need are never tested.
Escalate to the system owner when a hop fails because a process never captures the link; no graph work fixes that. Escalate to privacy or security owners when a fix would mean showing the agent records it cannot see today, or keeping links to records that retention rules say should go.
Sourcing workflow records with their links intact
Links are easily lost between the live system and a delivered dataset, because exports run table by table and de-identification can break join keys. A buyer who wants agents to work across a workflow should ask for the join keys, association tables and change history alongside the records.
When a buyer asks SourceX for workflow records, the search across U.S. businesses starts from the request form's scope, volume, language, format, timing and licensing fields, and listing the audit's hops there is a precise way to state the format. The business holding the records signs off on that scope and its terms before a single linked record is delivered.
Frequently asked questions
What is a missing link in a context graph?
It is a relationship a decision needs that the graph cannot traverse for a given case, such as a ticket that never references its order in a structured field. The records on both sides may exist and look complete. The link may be unrecorded, hidden in text, lost in export, invisible to the agent's identity or stale at decision time.
How is path completeness different from a linkage rate?
A linkage rate measures one relationship type across a dataset, such as the share of tickets linked to a customer. Path completeness measures whether every hop one decision needs is present for the same case. Because a decision fails if any hop fails, path completeness is usually lower than the lowest single linkage rate.
Should links extracted from free text count as complete?
Count them, but separately. Report path completeness with recorded links only and with extracted links included, and keep the source text span for every extracted link so a reviewer can check it. An agent should be able to see which kind of link it is relying on.
How many cases does a graph completeness audit need?
Enough to estimate the share of complete paths with useful precision and to include the case types where links break. A few dozen to a hundred closed cases from one workflow, with an extra stratum of escalated or reopened cases, is a practical start. Repeat the audit after each fix with the same sampling rule.
Can an agent work around missing links by searching?
Sometimes, but search substitutes a guess for a recorded link. A search for the customer's orders often returns the most recent one, not the one the ticket was about. If search is allowed as a fallback, evaluate it as an inferred link with a measured error rate.
Does a temporal graph solve missing links?
No. Validity windows help an agent avoid using links that were superseded, but they cannot create a link that was never captured. Temporal modeling and completeness auditing address different failures and are best used together.
Sources
- SHACL lets a property shape follow a sequence path, sets a minimum number of values with sh:minCount and reports each failure with its focus node and result path. Source
- PROV-O defines wasDerivedFrom, wasAttributedTo and generatedAtTime for recording where an entity came from, who it is ascribed to and when it was created. Source
- Graphiti describes facts with validity windows that are invalidated rather than deleted when information changes, each traced to the episodes that produced it. Source
- Zendesk tickets link an incident to its problem through problem_id, and follow-ups are created from a closed ticket through via_followup_source_id. Source
- HubSpot associations can carry labels, including paired labels such as parent and child, and one associated company can be marked primary. Source
- RFC 5322 uses In-Reply-To to identify the message being replied to and References to identify a thread of conversation. Source
Tell us what data you need
Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.