Structured metadata
The workflow returned defined fields rather than an unstructured paragraph. This makes the result usable by a database, CMS, review interface or publishing pipeline.
OJC LABS CASE STUDY · AI CONTENT ENRICHMENT
Supabase AI content enrichment with validation and semantic internal linking.
Publishing more content creates a second problem: every new page needs metadata, classification, related content and relevant internal links. When that work remains manual, it becomes inconsistent. When it is handed blindly to an LLM, it becomes confidently inconsistent at scale.
OJC Labs tested a Supabase AI enrichment engine as part of an AI content automation system, combining structured data, custom Python, multi-agent logic, validation and pgvector to generate contextually relevant internal-link candidates.
The validated workflow processed 200 records per hour. The point was not simply to make AI move quickly. The Lab test proved that speed could be combined with structured-output checks and semantic similarity rather than matching pages because they happened to repeat the same keyword.
01Content recordsrecord→
02Enrichmentenrich→
03Validationvalidate→
04Semantic matchembed→
05Approvalretrieve→
06CMSreview
200 records per hourin the validated processing benchmark
Structured outputs checkedbefore records moved forward
Embedding similarity + Supabase pgvectorused for contextual matching
Taxonomy and categorizationgenerated as structured data
Semantic internal linksselected by meaning rather than exact-keyword overlap
The content-scale problem
Publishing is only the beginning of content operations
As a website grows, each new page adds work far beyond writing:
- assign the correct category and topic;
- generate and verify metadata;
- connect the page to relevant commercial and educational content;
- avoid linking to the same convenient destinations every time;
- keep language and content-type rules intact;
- prevent orphan pages;
- update older records when the taxonomy changes;
- validate that enrichment is complete before publication.
Without a content operating system, teams either perform this work manually or skip it. The result is an archive full of individually decent pages that search engines and users struggle to navigate as one coherent body of expertise.
OJC Labs built the enrichment engine to test whether this layer could be industrialized without giving a model unsupervised control over site architecture.
Why OJC Labs tested it before selling it
Semantic matching sounds simple until irrelevant links start publishing themselves
The word “AI” makes internal linking sound magically solved. It is not.
A production system has to determine which records are eligible, which fields are mandatory, what counts as a related page and what happens when the model returns the wrong structure. It must also keep enrichment data separate from the original content so results can be reviewed, updated or rejected.
The Lab validation focused on four commercial risks:
- Throughput: can the workflow handle a meaningful content queue?
- Consistency: can every record return the same required data structure?
- Relevance: can linking candidates be selected from semantic meaning instead of loose keyword overlap?
- Control: can invalid outputs be stopped before they affect the content system?
The 200-record-per-hour benchmark demonstrated processing capacity. The workflow evidence demonstrated that categorization, structured validation and semantic matching operated as connected parts of one engine.
WORKFLOW EXECUTION OVERVIEW
Complete enrichment orchestration
What the validated engine produced
Structured enrichment that downstream systems can use
The engine transformed content records into operational data for search, navigation, editorial review and future publishing workflows.
STRUCTURED OUTPUT AND VALIDATION
One enriched record with visible validation state
The outcome is not a pile of generated suggestions. It is an enrichment layer that other systems can query, review and use.
Taxonomy and categorization
Content was classified into a controlled topic structure. A taxonomy is the agreed organization of categories, topics and relationships used across the content operation. It prevents every author or model from inventing a new label for the same subject.
Semantic representations
The system used embeddings, numerical representations of meaning, to compare content records. Supabase documents both the pgvector extension and semantic search, which stored and queried those vectors so related pages could be identified even when they did not repeat identical phrases.
Internal-link candidates
Embedding similarity was used to identify contextually related destinations. This supports semantic internal linking, where links are proposed because two pages are meaningfully connected, not because a keyword happened to appear in both.
Validation status
Structured outputs were checked before the record progressed. This makes failure visible and gives the wider system a reliable state to act on instead of assuming every AI response is complete. Supabase's automatic-embeddings pattern provides an external reference for queueing and synchronizing embedding work.
Why Supabase was useful in the validated stack
One data layer for content records and vector relationships
Supabase provided a Postgres foundation with pgvector support. In plain English, that allowed the workflow to keep structured content fields and semantic representations in one queryable data environment.
For the Lab test, this supported:
- structured content records;
- enrichment fields and processing states;
- vector embeddings;
- similarity queries for related content;
- downstream access for Python and workflow logic.
OJC Labs is not selling “Supabase installation” through this case study. The service is the complete enrichment operation: data model, processing logic, model controls, validation, semantic retrieval and integration with the client's CMS or content workflow.
Supabase is one validated implementation choice. The final architecture should still follow the client's data volume, security, hosting, editorial and integration requirements, including documented Row Level Security controls for the Data API.
SUPABASE CONTENT AND VECTOR FIELDS
Redacted content and vector fields
RANKED SEMANTIC LINK CANDIDATESCANDIDATE, NOT AUTO-PUBLISHED
- candidate
- approved
- rejected
- published
Source record and ranked related-page candidates
Batch validation video coming soon
From keywords to meaning
Better matching creates a better content network
Traditional automatic linking often relies on exact phrases. That creates obvious problems:
- several unrelated pages can repeat the same broad keyword;
- two highly relevant pages may use different terminology;
- the same high-authority destination can be recommended repeatedly;
- anchors can become mechanical and over-optimized;
- multilingual and specialist content can be misunderstood.
Semantic similarity evaluates the meaning represented by the content. It can therefore surface relationships that exact matching misses. Where exact terms and meaning both matter, Supabase also documents hybrid search, which combines keyword and semantic retrieval.
That does not mean every high-similarity result should be published automatically. A client system may also need rules for locale, canonical destination, status code, content type, maximum link count, commercial priority and editorial approval.
OJC Labs designs those controls around the site. The model can identify meaning. The operating system still decides what is allowed. Our guide to internal-linking strategies that work explains the editorial and SEO layer around the matching engine.
Validated results
The engine proved capacity and controlled output
- Processing speed reached 200 content records per hour in the validated workflow.
- The engine generated structured metadata, taxonomy and categorization data.
- Structured outputs were checked before downstream use.
- Custom Python and multi-agent logic coordinated the enrichment stages.
- Embedding similarity and Supabase pgvector supported semantic related-page retrieval.
- Internal-link generation moved beyond exact-keyword matching.
- The working flow provided a technical proof base for a client-ready AI content-enrichment service.
The available evidence does not support claims about ranking increases, traffic growth, revenue or a specific reduction in orphan pages. Those outcomes should be measured in the client's analytics and crawl data after deployment, not retrofitted onto a Lab benchmark because the marketing department felt optimistic.
What OJC Labs can build for a client
AI enrichment and internal-linking automation as a service
OJC Labs can design and integrate:
- Content-corpus and sitemap ingestion;
- Metadata and taxonomy enrichment;
- Structured-output schemas and validation;
- Vector-embedding generation and storage;
- Semantic or hybrid content retrieval;
- Internal-link candidate generation;
- Locale, canonical and content-type restrictions;
- Anchor-text and repetition controls;
- Approval queues for editors or SEO teams;
- CMS or database writeback;
- Crawl validation after publication;
- Monitoring, retries and audit logs;
- Performance reporting across processed, accepted and rejected suggestions.
The service can support editorial websites, multilingual content libraries, ecommerce education hubs, media archives, documentation portals and companies publishing large volumes of expert content. It can connect directly to wider SEO and indexing systems and to an existing CMS by automating publishing with APIs.
What should remain human
Automation can prioritize. Editorial teams still own meaning and risk.
A production implementation can automate high-confidence, low-risk enrichment and route ambiguous suggestions for review. Human control remains especially important when:
- A link affects a regulated or legally sensitive page;
- The candidate changes commercial priority;
- The content uses specialist terminology;
- Translated pages have different available destinations;
- An anchor could imply a claim not supported by the destination;
- A content owner is reorganizing the taxonomy.
The correct approval model depends on publishing volume and consequence. OJC Labs does not confuse autonomy with the absence of accountability.
Frequently asked questions
AI content enrichment, explained
What is AI content enrichment?
AI content enrichment adds structured information to an existing content record. This can include summaries, metadata, categories, entities, topics, related pages, embeddings and quality checks. The purpose is to make content easier to organize, retrieve, publish and connect—as demonstrated in the AI is Mid media-system case study. Teams designing the wider architecture can also review building an AI content pipeline.
What is semantic internal linking?
Semantic internal linking uses the meaning and context of pages to identify relevant relationships. It can connect content that discusses the same concept with different wording. Additional business, SEO and editorial rules should still decide whether a candidate link is eligible for publication.
What is pgvector?
pgvector is a Postgres extension for storing and comparing vector embeddings. An embedding is a numerical representation that allows systems to compare the semantic similarity of text or other data.
Why use Supabase for an enrichment engine?
Supabase combines Postgres with APIs and vector capabilities in one platform. It can be useful when structured content, processing state and embeddings need to remain queryable by the wider workflow. OJC Labs confirms the final architecture against each client's requirements.
Will the system publish internal links automatically?
It can, but it does not have to. OJC Labs can configure automatic publication for high-confidence candidates, human approval for all suggestions, or a hybrid model. Destination health, locale, canonical and content rules should be validated before any writeback.
Does automated internal linking guarantee higher rankings?
No. Relevant internal links can improve discovery, crawl paths, site structure and authority distribution, but rankings also depend on content quality, external authority, technical SEO, intent, competition and many other factors.
Can this enrich existing content?
Yes. The engine can process new or historical records, provided the source data and destination systems are accessible and the enrichment schema is defined.