Alexander Kihm leads POMA AI, a Berlin-based context intelligence laboratory focused on retrieval infrastructure for AI systems. His path to founding the company reveals a pattern of tackling unglamorous infrastructure problems: after training as an industrial engineer at Karlsruhe Institute of Technology and earning a PhD in big data econometrics at the German Aerospace Center—where he worked on the world's first mass-market plug-in hybrid car—he founded Fairr, a fintech that modernised Germany's pension system before being acquired by Raisin, and Advo-Assist, which grew into Germany's largest digital platform for lawyers.
What drew Kihm to artificial intelligence was a deceptively simple observation: despite dramatic monthly improvements in model capabilities, companies deploying these systems on their own documents saw no corresponding gains in answer quality. "When you trace those failures back, the model is rarely the culprit," he explains. "The problem sits upstream, in how documents are cut up, described and stored before the model ever sees them. That plumbing is where I decided to spend my time."
The Retrieval Problem
POMA's core offering is an ingestion engine that transforms lengthy, unstructured documents—reports, manuals, contracts, scanned forms—into retrievable pieces that retain their hierarchical position within the source material. As Kihm describes it, a figure on page 40 carries little meaning in isolation; it becomes actionable once the system knows it belongs under "Q3 payroll, Berlin office." His role encompasses founder, CEO, CTO and chief scientist, responsibilities that keep him involved in both company operations and the research behind the benchmarks POMA publishes openly.
The retrieval step itself fails far more frequently than most organisations realise. When an AI assistant must locate the correct passage, table or figure from a company's document collection to answer a user's question, this foundational task often determines whether the final response will be accurate or misleading.
Why Images Matter
In numerous sectors, the image is the document. A maintenance manual conveys a valve assembly through an exploded diagram rather than prose. Architectural blueprints encode dimensions, room identifications and structural notes that exist nowhere else. Bank forms captured as scans contain account holder birthdates as pixels, not searchable text. When AI systems restrict themselves to text-only search, all this visual information becomes invisible to the assistant, leaving employees to conclude the tool "doesn't know" what they need and reverting to manual PDF scrolling.
Kihm emphasises that handling images correctly transforms an AI assistant from "a nice demo on clean prose into something an engineer, architect or clinician will actually rely on."
The Vector Compression Problem
Most image retrieval systems convert each image into a single vector—a compressed numerical representation of the entire picture. This approach captures broad themes effectively but performs poorly on specifics, much like viewing a page through frosted glass where you recognise it's a financial table but cannot read individual figures.
The classic "red car" problem illustrates the limitation: searching for a red car returns red objects that aren't cars and cars that aren't red, because the system matched on loose conceptual similarity rather than actual visual content. Exact numerical details expose the weakness further. In POMA's testing, when searching for a specific payroll figure of 852.79, image-based search ranked a US Treasury bulletin page first and the actual payroll page fifth—both appeared as "a page of numbers" to the image model, making them near neighbours despite one being entirely wrong. When the same pages were searched using detailed text descriptions instead, both correct pages ranked first.
The difference between fifth place and first place matters enormously: an AI agent reading only the top result will confidently return the wrong number.
Context as Meaning
Context fundamentally transforms how images are understood. Consider a zoomed X-ray of bone: without context it remains an abstract pattern of black, grey and white; even trained professionals require knowing which bone, which viewing angle and the reason for imaging. The caption—"left distal radius, lateral view, two weeks post-fracture"—and surrounding clinical text provide the image with meaning.
The same principle applies to charts missing axis labels or floor plans without title blocks indicating building and floor. Image pixels show what something looks like; context reveals what it is and why it matters. Without context, retrieval either casts too wide a net (every X-ray resembles every other X-ray) or too narrow (the needed image never matches the query).
Building Better Descriptions
Detailed visual descriptions transform fuzzy images into precise, searchable language. If search functions as pointing at content, vague descriptions cast a broad beam while quality descriptions function as a laser targeting exactly the needed figure.
According to Kihm, genuinely useful descriptions contain four essential components:
- Concrete representation of what appears: "exploded view of a hydraulic pump with seven numbered components" rather than "a technical drawing"
- Every text element within the image, transcribed precisely: labels, part numbers, figures, dates. While image models capture some text, they do so loosely, so exact label searches may fail without transcription
- The original caption, typically the creator's own explanation of the image's purpose
- The image's location: which document, chapter and heading contain it. Most systems omit this final element, yet it frequently distinguishes between otherwise identical-looking pages
Why Early Intervention Matters
Information discarded during ingestion cannot be recovered downstream. Conventional pipelines begin by dividing documents into fixed-size chunks and extracting images as standalone files. At this moment, figures lose their captions, tables lose their headers, and paragraphs lose the headings explaining their context. Everything that follows—rerankers, larger models, sophisticated prompts—attempts reconstructing meaning from fragments, which explains why many retrieval improvement efforts disappoint: they intervene too late.
POMA's context tree preserves the document's structural framework. Every content piece, including images, carries its hierarchical path: document, chapter, section, sub-heading. Rather than storing merely "a table containing 35," the system stores "35, under International Capital Movements, in the 2023 Treasury Bulletin." This path distinguishes between similar-looking items.
Testing across over 21,000 chunks demonstrated the impact: adding hierarchical context moved the correct answer from a median rank of approximately 1,780 to approximately 15—a roughly 119-fold improvement. The prevailing assumption that extra context "dilutes" a search vector proved incorrect in practice; context actually disambiguates it.
Smaller Models, Better Results
Generating visual descriptions for retrieval represents a narrow, repetitive task executed at massive scale. A company might possess millions of pages, each containing multiple figures. Running a frontier general-purpose model across all of this becomes prohibitively slow and expensive, and organisations pay for capabilities—poetry, coding, general conversation—that this specific task never requires.
Quality considerations also favour specialised models. The ideal model receives the "Goldilocks" amount of information: the image, its caption, nearby text and its position in the context tree. Too little information and it guesses; too much (like a full 300-page document) and it becomes distracted while costs soar. A smaller model trained specifically for this task, receiving exactly that information slice, operates faster, cheaper and more consistently. It can also run on a company's own hardware—critical for medical, legal or engineering documents that cannot leave the premises.
POMA's on-premises product exemplifies this approach. It operates a single small model of approximately 4 billion parameters on one standard GPU within the customer's infrastructure, completely disconnected from the internet, ensuring documents never leave the building. Lightweight specialisations train atop this foundation for specific tasks, including one dedicated to describing images, charts and figures in documents. In POMA's testing, their small specialised model describes images at least as well as general-purpose models up to a hundred times larger. Because it remains small, the company can further adapt it to a client's specific domain and documents—whether engineering drawings, medical imagery or financial forms—making it even more effective. "You don't need the biggest model; you need the right one for the job," Kihm states.
Unlocking Existing Data
Most organisations already possess answers to a substantial portion of their questions. These answers lie buried in manuals, drawings, scanned forms and presentations that nobody can search effectively. Better retrieval doesn't generate new data; it makes existing data accessible—usually the cheaper and faster path to value.
This becomes even more critical with AI agents, which search differently than humans. A person executes one search, reviews results and applies judgment. An agent may launch dozens of queries in succession and act on whatever appears first. If the correct figure ranks fifth, a human might spot it; an agent typically won't.
For companies seeking to extract more value from information they already possess, Kihm recommends three priorities:
- Audit what remains locked in images. Diagrams, scans and charts frequently contain the most valuable yet least searchable information
- Fix ingestion before purchasing a larger model. Preserve structure, captions and context at the beginning; you cannot restore them later
- Test retrieval independently. Pose real questions with known answers and verify whether the correct source ranks first, not merely whether the final answer sounds reasonable
Organisations that execute this correctly discover their AI agents suddenly "know" substantially more—not because the underlying model became smarter, but because it can finally access what was present all along.
Source: TechRound
