Building a Precision Search Engine Without Vectors: A Practical Approach for SAP Commerce Cloud

Introduction

Search technology has entered a new era with the rise of vector embeddings and neural retrieval models. While these approaches offer powerful capabilities, they also introduce complexity, resource overhead, and opaque debugging challenges. At HybrisArchitect.com, we recently implemented a multilingual, natural language–friendly search engine that focuses on PDF document retrieval, integration with SAP Commerce OCC APIs, and a company’s website FAQs. Interestingly, we achieved strong performance using traditional Solr-based search without vector embeddings.

One of the impetuses for this article was noticing that many websites providing product specifications in PDF format do not make these files searchable. To address this observation, we downloaded over 300 different PDF product specification sheets to build our search index. The results have been encouraging, demonstrating that structured, deterministic search can deliver precise outcomes across large document collections.

This article explores why a deterministic, vectorless approach can be not only sufficient but also highly effective for many enterprise commerce use cases.

Why Vectorless Search?

1. Deterministic Precision

In regulated and commerce-heavy environments, reproducibility and explainability matter. With a vectorless Solr implementation:

• Search results are easier to debug and explain to business teams.
• Queries produce consistent, predictable outcomes.
• Ranking logic can be tuned transparently with BM25, field boosts, and rerankers.

2. Performance and Simplicity

• No GPU overhead or embedding pipelines to maintain.
• Faster indexing, especially for large PDF corpora.
• Less infrastructure complexity means easier scaling and maintenance.

3. Competitive Accuracy

Through experimentation, we found that:

• Hybrid search (vector + BM25) often performed similarly, but the added complexity rarely justified the gains.
• Reranker + BM25 strategies consistently matched or outperformed hybrid vector setups for deterministic document queries.
• Business users preferred the clarity of “why” a result ranked where it did.

Case Study: Identify Specific Search Term in Closely Related Documents

To conduct the search, we connected the data sources to a chatbot and queried the term: ‘What are Sheathing installation instructions?’

The query was executed over 300 construction-related PDF product specification sheets, entirely without embeddings and without fine-tuning. Despite this, the search engine delivered precise, formatted results that were easy to follow and included proper citations.

Chatbot Output Example:

As you can see, the returned message format was improved for readability and user experience. Results were presented with clarity, making them easier to interpret, and the sources were cited at the bottom of the response.

Original PDF Source:

Below is the original PDF source, which contains the installation directions in a slightly hard-to-follow format. The chatbot reformatted the response in a much more intuitive way while maintaining the integrity of the original source.

This case study highlights that deterministic Solr search, combined with thoughtful output formatting, can deliver production-grade accuracy for specialized document queries.

Lessons Learned

1. Rerankers are underrated: Applying a lightweight reranker on top of BM25 often bridges the gap that vectors were supposed to fill.
2. Explainability wins trust: Business stakeholders trust search more when they can see why results surfaced.
3. Less can be more: Avoiding embedding generation pipelines reduces infrastructure cost, complexity, and operational risk.
4. Multilingual doesn’t require vectors: With proper Solr configuration (language-specific analyzers), we achieved strong coverage across languages.

Conclusion

While vectors and hybrid retrieval approaches dominate current discussions, our experience demonstrates that vectorless Solr search remains a powerful, precise, and enterprise-ready option. For scenarios where explainability, performance, and deterministic results are key—such as SAP Commerce product catalogs, PDF-heavy documentation, and FAQs—traditional approaches still shine.

At HybrisArchitect.com, we’ll continue exploring innovative but pragmatic search strategies that deliver measurable business value without unnecessary complexity.

If you would like to see a demo or explore how this approach could work within your organization, please reach out to us at info@localhost.