How we index images for RAG

TEXT START: Kapa builds AI assistants that answer questions from technical documentation.

THE DISSECTION

This is a technical engineering blog post describing how Kapa solved the economic problem of processing images in RAG pipelines at scale. The post is structured as engineering documentation: problem identification, failed approaches, successful solution, implementation details, measured results. The veneer is pure craft and technical rigor. The function is recruitment and customer acquisition dressed as knowledge sharing.

THE CORE FALLACY (DT LENS)

The article optimizes within a paradigm it never questions: that automating the answering of technical questions is a product to be sold, not a displacement to be reckoned with. The DT thesis says AI severs the employment-wage-consumption circuit by automating cognitive work. This post is a case study in exactly that severance, narrated as infrastructure improvement.

The business model is: replace the human support worker who answers "where do I click?" by building an AI that reads your documentation, sees your screenshots, and answers. The metric of success is explicit: users who "self-serve" instead of "open a ticket." Fewer tickets = fewer support jobs. The optimization is real; the displacement is the point.

HIDDEN ASSUMPTIONS

1. The knowledge work being automated is simply tedious, not structurally necessary — "configuring X" questions are treated as grunt work to be automated, not as the kind of specific, contextual human judgment that resisted automation until now.
2. Higher accuracy at lower cost is unambiguously good — The post measures "statistically significant improvement" as a win without asking what happens to the humans whose tickets are now unnecessary.
3. The paradigm is stable — This is pure within-system optimization. No acknowledgment that the RAG pipeline being optimized is itself a transient artifact of the transition between human cognitive labor and AI cognitive labor.
4. Indexing cost is the bottleneck — They chose "describe once" because query-time vision is expensive. This ignores that compute costs compress over time; their optimization may be economically irrelevant within two years as vision inference costs drop.

SOCIAL FUNCTION

Transition management. The post tells a story of engineers solving a hard problem well, providing cover for the reality that the product being built automates human cognitive labor. The technical depth signals credibility; the measured results signal success; the absence of any discussion of who loses signals that there's nothing to discuss. This is the ideological work that makes cognitive automation feel like infrastructure improvement rather than labor displacement.

THE VERDICT

This is a well-engineered solution to a real technical problem. It is also a machine for producing support worker unemployment at scale, narrated as engineering best practices. The post itself notes the value proposition: answers users can act on "without hunting for the setting" — meaning fewer humans needed to point users at settings. The DT lens does not judge the engineering; it notes that this engineering is a blade in the circuit between productive labor and consumption.