Vector Search & Embeddings

A recurring research example

Consider a concrete scenario: a novelty search for a "cooling mechanism for a battery unit using microchannels". The search in a standard patent database begins with a classical Boolean query (battery AND cooling AND microchannel), extended by synonyms and the matching CPC class. The result is a solid hit list, accompanied by a lingering doubt that relevant citations were missed.

The doubt is justified. The following sections show where classical search reaches its limits, how semantic search, hybrid search, and retrieval-augmented generation (RAG) help, and where these technologies also have weaknesses that must be understood for professional use.

The limits of classical keyword search

Boolean logic works well where the terminology of all relevant documents is predictable. For international, interdisciplinary technologies, that is rarely the case. Five typical gaps:

At least one of these gaps opens in almost every larger search.

Semantic search: meaning instead of letters

Instead of searching for words, semantic search looks for meaning. Each text fragment (a claim, a paragraph, a published document) is translated into a position on a kind of map of technical concepts. Texts with similar content end up close together, regardless of the terminology used.

In the research example, semantic search finds the three citations that keyword search missed: the EP application with different wording, the Japanese filing, and the microfluidics publication from a neighboring domain.

The practical benefit lies in a different hit list, not a longer one. Documents become visible that keyword search would have overlooked. Conversely, documents are filtered out that do contain the keywords but are not substantively on point.

Where semantic search fails: hybrid search

Semantic search is powerful, but it has precise blind spots, and those blind spots sit exactly where patent search demands precision:

Hybrid search combines both: keyword exactness for strings plus meaning-based similarity for concepts. Both ranked lists are fused into one result list. For patent search, that is mandatory, not optional: pure semantic search loses strings, pure keyword search loses meaning.

RAG: the research assistant with file access

A sorted hit list is a beginning, but not an answer. The next level is retrieval-augmented generation. The language model receives the retrieved citations as context and answers questions on that basis, not from its training knowledge, but from the documents specifically handed to it.

The analogy will be familiar to patent practitioners from their own training: RAG works like the open-book principle of the European Qualifying Examination (EQE). The system may look things up, but its answer must be supported by the provided materials. The legal assessment (novelty, inventive step, sufficient disclosure) remains the task of the human, not the model.

Concrete applications in the research example:

• "Summarize the relevance of the top-10 hits against claim 1."
• "Which features of claim 1 of EP 3 456 789 cover the independent claim of the present application?"
• "From 47 office actions over the past two years concerning H01M 10/61, extract the examiner's typical objections."

The retrieval quality of such systems is empirically supported: Ding et al. (arXiv:2508.14064, 2025) report precision of 80.5 % and recall of 92.1 % for a RAG-based patent retrieval system. That lies well above approaches without retrieval integration.

For practical implementation with local, privacy-preserving tooling, the RAG workshop guide describes a complete pipeline with Ollama and ChromaDB. For autonomous research systems that carry out multiple RAG steps on their own, see AI agents and isolated context.

Limits of RAG, honestly assessed

RAG does not solve every problem either. Seven limits that must be understood for professional use:

Visible only inside the index

RAG finds nothing that has not been indexed. A complete prior art search is not replaced by the system; it accelerates initial review and result structuring.

Retrieval quality is the bottleneck

If poor chunks land in the context, even the best prompt yields a weak answer. Grounding is only as good as the underlying retrieval.

Hallucinations despite grounding

Rare but real: the model invents paragraph numbers, citations, or references that do not appear in the supplied chunks. Every cited location must be checked against the original document. A California attorney was fined USD 10,000 in 2024 after 21 of 23 AI-generated citations in his appeal brief were identified as fabricated. The risk is not theoretical.

Chunking breaks context

The underappreciated problem of patent RAG. When patent documents are split into text segments so individual pieces can be retrieved, structural connections are lost. Three typical failures:

• Prior art contamination: paragraph 0012 describes the invention and is retrieved as a relevant chunk. The next paragraph, 0013, contains the transition sentence "In the following, the prior art is described" and is dropped as not sufficiently relevant. Paragraph 0014 describes a prior art feature without repeating the term, and is also retrieved as a relevant chunk. The language model sees the two chunks side by side and reads the feature from 0014 as part of the invention. The decisive information sat in the skipped paragraph between them.
• Negation inversion: the sentence "the device does not contain a cooling unit" is split during chunking. If only "cooling unit" survives, the meaning flips.
• Loss of claim dependency: a dependent claim ("according to claim 1, characterized in that…") is retrieved in isolation. Without the reference to the independent claim, it reads like standalone teaching. For a novelty or inventive-step analysis, that is a structural error.

This list is not exhaustive. Beyond the three classic cases, chunks routinely break on term definitions that sit in a non-retrieved paragraph ("In the context of this disclosure, …"), on reference-number chains distributed across paragraphs (a housing (1), with a mount (2), holding a heat sink (3), which carries the microchannels (4)), and on functional, mechanical, spatial, temporal, or causal relationships carried by grammatical structures such as "wherein", "which", or anaphoric pronouns. These relationships have no signal words a chunker could detect.

Section-aware chunking helps where section boundaries are the actual problem, as in the prior art contamination case above. Within a section, it has no handle. Parent-document retrieval (pulling neighbouring chunks alongside) and definition extraction (pre-processing terms and reference numbers) cover further cases, but no single mitigation catches every failure mode. For patent RAG, at least one such measure is required; naive chunking is systematically dangerous. The pragmatic consequence: the original document remains the final authority, especially where claim features or their interpretation are disputed.

Drawings are invisible

Pure text embeddings ignore figures and reference-sign links. For many patents (particularly in mechanics and electrical engineering), this is a real loss. Multimodal models are improving, but are not yet production standard.

Imprecise citations

The model produces "per EP 3 456 789, paragraph 23". The paragraph number may be shifted, the citation distorted. Citations must always be verified against the original document.

Index aging

Documents published after the last index build are missing. For ongoing matters, at least weekly updates are advisable.

Six rules for professional use

The rules of work follow from the limits. None are new, but they apply more sharply to AI-supported research than to manual work:

1. Always hybrid search, never pure semantics. For patent search there is no good reason to drop the keyword component. Application numbers, CPC codes, and applicant names must match exactly.
2. Check RAG answers against the original chunks. Every citation verified, never adopted blindly. For any substantive conclusion: back to the original text.
3. RAG for scoping, not for the opinion. The system accelerates initial review. The legal assessment stays with the human. The practical test: can the AI result be verified faster than the task can be done directly? If not, there is illusory efficiency with loss of control.
4. Keep an audit trail. Which chunk led to which conclusion? For two reasons: a second opinion within the firm, and liability documentation if results are later challenged.
5. Client data local only. Cloud AI^* is subject to strict professional requirements under German and European law (§ 39c PAO, § 203 StGB). For confidential invention information, local operation is the clean solution, for instance with Qdrant or ChromaDB plus an open-source embedding model. This has been practically viable since Ollama added native Anthropic API support in January 2026. Details in the RAG workshop guide.

Outlook

Development is moving in three directions particularly relevant to patent practice: multimodal models that can evaluate drawings and reference signs; agent-based systems that carry out research tasks autonomously across multiple steps (see AI agents and isolated context); and better explainability, making it traceable why a hit was ranked as relevant. Each of these shifts the boundary of sensible AI use. None changes the fundamental rule: the system proposes, the patent practitioner decides.

Sources

1. Goebel, "Künstliche Intelligenz im Patentwesen – Revolution oder Evolution?", GRUR Patent 2025, 592 (in German; beck-online)
2. Helmers et al., "Automating the search for a patent's prior art with a full text similarity search", PLoS ONE 14(3), 2019
3. Lewis et al., "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks", NeurIPS 33, 2020
4. Ding et al., "An automatic patent literature retrieval system based on LLM-RAG", arXiv:2508.14064, 2025
5. Anthropic, "Contextual Retrieval", 2024 (anthropic.com)
6. Liu et al., "Lost in the Middle: How Language Models Use Long Contexts", TACL 12, 2024
7. Google Skills, "Vector Search and Embeddings" (course): skills.google
8. Vertex AI Documentation, "About hybrid search": cloud.google.com