If your company searches for contracts, procedures or internal emails and the usual answer is «I can't find it» or «I don't know where it is anymore», the problem is not filing: it's search technology. Classic text engines — based on exact keywords — fail when the user writes in natural language or when the document uses different synonyms than those that appear in the query. A vector index solves exactly that: it converts each text fragment into a mathematical representation (a high-dimensional vector) that captures meaning, not just words. The result is a search engine that understands «show me the maintenance contracts expiring this quarter» even if the document says «periodic review» and «Q2».
But how much does it cost to build that in a Spanish SME with 20 to 200 employees? The honest answer is that it depends on several factors we develop below, but market ranges in 2026 move between €3,000 and €40,000 for the initial project, with monthly maintenance costs of between €80 and €800 once in production. This article breaks down those figures, explains what drives them and helps you calibrate what level you actually need.
What is a vector index and what problem does it solve in an SME?
A vector index is a database specialised in storing and querying embeddings: numerical representations of text (or images, or audio) generated by a language model. When a user submits a query, the system converts it into a vector as well and finds the document fragments closest in that mathematical space. This process is called semantic similarity search or, in a generative AI context, it forms the retrieval layer of a RAG (Retrieval-Augmented Generation) system.
In practice, what an operations director cares about is not the architecture but the outcome: the system reads all your PDFs, emails, spreadsheets and Word documents, indexes them once and then answers questions in natural language about that content, citing the exact fragment where it found the answer. For an SME with 5,000 internal documents, the difference between finding the right clause in 8 seconds versus 45 minutes is real and quantifiable.
Factors that determine the price
There is no catalogue price for this technology because the final cost depends on six variables that are worth understanding before asking for a quote:
1. Volume and format of the document corpus
The cost of generating the initial embeddings scales with the number of tokens (units of text). A corpus of 10,000 pages in well-structured PDFs is very different from 10,000 pages scanned as images that require prior OCR processing. Extracting text from image-format documents (scanned invoices, paper-signed contracts) can account for between 20% and 40% of the total project cost if the volume is high.
2. Choice of embedding model
Embeddings are generated by a language model. There are two paths here: use the API of an external provider (OpenAI text-embedding-3-large, Cohere Embed v3, Google text-embedding-004) or deploy an open-source model on your own server (BGE-M3, E5-mistral, nomic-embed-text). The first is faster to deploy and requires less infrastructure, but it means that documents leave your perimeter to be processed. The second keeps data within your network, but requires hardware and maintenance.
3. Vector index engine chosen
The most common options in SME-scale projects are Qdrant, Weaviate, Milvus, Chroma and pgvector (a PostgreSQL extension). Each has a different profile of operational cost, ease of management and scalability. For volumes of up to 500,000 vectors, pgvector on an existing PostgreSQL database may be sufficient and eliminates the need for an additional service. For millions of vectors or high-concurrency searches, specialised solutions such as Qdrant or Weaviate — also available in managed cloud mode — perform better.
4. Integration with the existing ecosystem
Vector search alone is useless if users cannot invoke it from where they work. The integration layer — SharePoint, Google Workspace, the ERP, an internal chatbot in Teams or Slack — can double the project cost if built from scratch, or be drastically reduced if a framework like LangChain or LlamaIndex is used, which already includes connectors for the most common platforms.
5. Deployment infrastructure
Does the solution live in the provider's cloud, on your own server, or on a dedicated server managed by Summum? The cloud option (AWS, Azure, Google Cloud) offers elasticity but generates a recurring cost for API calls and storage. The on-premise option eliminates the variable cost but requires an initial investment in hardware (a GPU for the model, although small embedding models work well on CPU) and in maintenance.
6. Maintenance and continuous ingestion
A company's documents are not static. Every week new contracts, emails and procedures are generated. The system needs an ingestion pipeline that detects new or modified documents, processes them and updates the index. This component — which seems minor — makes the difference between a solution that becomes stale in three months and one that remains useful long-term.
Orientative price ranges by scenario (2025–2026 market)
The following ranges reflect market prices for projects in Spain with specialist technical consultancy. They are not Summum's rates but references drawn from industry publications (Gartner, IDC, vendor reports from Qdrant and Weaviate) and direct market observation in similar projects.
| Scenario | Profile | Project cost | Monthly cost (production) |
|---|---|---|---|
| Basic pilot | Up to 5,000 docs, one source type (e.g. SharePoint), simple web interface, cloud | €3,000 – €7,000 | €80 – €200 |
| Standard SME solution | 10,000–50,000 docs, multiple sources (SharePoint + ERP + email), Teams/Slack integration, RAG with LLM | €8,000 – €18,000 | €200 – €500 |
| Solution with sensitive data (on-premise) | Same as standard but with open-source model deployed on own server; data does not leave the perimeter | €12,000 – €28,000 | €150 – €400 (own infra) |
| Multi-domain / mid-market | 100,000+ docs, multiple departments, role-based access control, continuous ingestion pipeline, managed SLA | €25,000 – €40,000 | €400 – €800 |
Note: cloud infrastructure costs (embedding API calls, index storage) vary by provider and volume. With OpenAI text-embedding-3-small, indexing one million tokens costs approximately $0.02 (rate published by OpenAI in 2025); for a corpus of 10,000 dense pages this amounts to less than $5 in the initial run.
What line items make up the budget?
A typical budget breakdown for an SME vector index project includes the following items. Percentages are orientative over total cost:
- Analysis and architecture design (10–15%): scope definition, technology selection, data governance plan.
- Document extraction and cleaning (15–30%): connectors to SharePoint, Google Drive, ERP or file server; OCR if scanned documents exist; chunking and pre-processing.
- Initial index generation and loading (5–15%): embedding model calls, loading into the vector engine, retrieval quality tests.
- Query layer and interface development (25–35%): search API, integration with Teams / internal web, prompt engineering if generative response (RAG) is included.
- Continuous ingestion pipeline (10–20%): change detection, incremental re-indexing, management of permissions inherited from the original source.
- Training and go-live (5–10%): documentation, training for key users, initial monitoring.
If you want to understand how this architecture fits into a broader augmented information retrieval strategy, our semantic search for companies service details the technology options and the step-by-step deployment process.
Comparison of vector engines for SMEs
The choice of engine is one of the decisions with the greatest impact on total cost. This table summarises the most common options in SME-scale projects in 2025–2026:
| Engine | Deployment model | Ideal for | Approximate cost (managed cloud) | Licence |
|---|---|---|---|---|
| pgvector | PostgreSQL extension (self-hosted or RDS/Supabase) | Corpus up to 500K vectors; team already uses PostgreSQL | €0 extra if you already have Postgres; RDS ~$30–80/month | Open source (PostgreSQL) |
| Qdrant | Docker / Qdrant Cloud | Millions of vectors, metadata filtering, high performance | From $25/month (Qdrant Cloud starter) | Open source (Apache 2.0) + SaaS |
| Weaviate | Docker / Weaviate Cloud | Multimodal, built-in embeddings, GraphQL | From $25/month (sandbox) | Open source (BSD) + SaaS |
| Chroma | Embedded Python or local server | Rapid prototyping, data science teams | €0 (self-hosted) / Chroma Cloud in beta | Open source (Apache 2.0) |
| Azure AI Search | Azure SaaS (with vector support since 2023) | Microsoft 365 / Copilot Studio ecosystem | From ~$70/month (Basic); varies by index and queries | Microsoft managed service |
The factor that most inflates project cost: the quality of source documents
Based on experience accumulated since 2007 working with companies of 10 to 250 employees in Castilla y León and the Canary Islands, the factor that most frequently pushes costs above the initial estimate is not the vector technology itself but the state of the source documents. There are three typical situations that inflate the project cost:
- Image documents without selectable text: signed and scanned contracts, PDF image invoices. These require an OCR step (Tesseract, Azure Document Intelligence, Google Document AI) that adds processing cost and, if scan quality is low, manual correction work.
- Archive fragmented across multiple locations: companies with documents in SharePoint, a local NAS server, an employee's personal Google Drive and email attachments. Each source requires a specific connector and a permissions policy.
- Absence of structured metadata: documents without a reliable creation date, without a category, without a department assigned. The vector index can search by content, but without metadata it cannot filter «only client contracts from the last year».
Dedicating a document audit phase before the technical project — even if it seems like an extra expense — usually reduces the total cost because it avoids surprises during the extraction phase and improves the quality of results from day one.
When does a vector index deliver ROI in an SME?
The question most directors ask before approving the budget is legitimate: when will it pay back? Return on investment in this type of project comes mainly from three sources:
- Reduction in search time: if an employee spends an average of 1.5 hours per week searching for documents (a conservative estimate according to the IDC study «The Hidden Cost of Information Retrieval», 2023), and the vector index reduces that time by 70%, the annual saving per employee is around 50 hours. At a cost of €25/hour, that is €1,250 per employee per year.
- Reduction of errors from using outdated versions: the index can be configured to always show the most recent version of a procedure or contract, reducing the risk of working with outdated documents.
- Acceleration of new employee onboarding: guided access to the internal document base shortens the time to productivity for new hires.
For a company of 30 employees, with a saving of €800 per employee per year (a prudent estimate), the total annual ROI would be €24,000. A €12,000 project pays for itself in less than six months. These figures vary greatly by sector and usage profile, but they illustrate that the profitability threshold is achievable even in small SMEs.
Cloud or on-premise? The GDPR factor
When the indexed documents contain personal data — employee contracts, client files, medical reports in the case of clinics — the option of sending them to an external API to generate embeddings requires a prior legal assessment. The General Data Protection Regulation (GDPR) requires that any transfer of personal data to an external provider be covered by a data processing agreement (DPA) and, if the server is outside the EU, by additional international transfer mechanisms.
OpenAI and Cohere offer DPAs and have infrastructure in the EU (Ireland, the Netherlands), which simplifies management. However, for sectors with particularly sensitive data — healthcare, legal, tax advisory — many companies prefer the on-premise option: an open-source embedding model (such as BGE-M3 or nomic-embed-text) deployed on their own server, with the vector index also local. In that scenario no data leaves the corporate perimeter, at the cost of greater technical deployment complexity and a slightly higher initial cost.
If you have doubts about how to structure data governance in a semantic search project, our semantic search for companies team will evaluate with you the most appropriate scenario based on your data and your sector.
Frequently asked questions
Can I deploy a vector index without having any developer on staff?
Yes. Most vector index projects for SMEs are executed entirely by the external consultancy. The client company does not need prior technical knowledge: it provides access to its document systems (SharePoint, file server, Google Drive) and defines what types of queries it wants to perform. The consultant designs the architecture, deploys it, integrates it with the tools the employees already use (Teams, internal web) and trains the team. It is common for an internal technical contact — even the part-time IT manager — to participate in validations, but it is not essential for the deployment.
How long does it take to have the solution running?
A pilot on a bounded corpus (up to 5,000 documents, a single source) can be operational in 3–6 weeks. A complete solution for a medium SME (multiple sources, Teams integration, continuous ingestion pipeline) requires between 2 and 4 months. The greatest time is consumed not by the vector technology itself but by extracting and cleaning the source documents and integrating with the client's working environment.
What is the difference between vector search and a full RAG system?
Vector search is the retrieval layer: it finds the document fragments most relevant to a query. RAG (Retrieval-Augmented Generation) adds a language model on top that synthesises those fragments into a prose response, with citations. For a corporate search engine where the user wants to see the original document, vector search alone may be sufficient and more cost-effective. If the goal is an assistant that answers questions in natural language («what is our returns policy for clients with more than €5,000 in annual purchases?»), then the full RAG system is needed, which also incorporates the cost of the generative model (LLM).
Will the system become obsolete if my company generates many new documents?
No, if a continuous ingestion pipeline is implemented from the start. This component monitors the source folders or systems, detects new or modified documents, processes them and updates the index automatically — normally in nightly cycles or near real-time depending on criticality. Without this pipeline, the index ages and loses relevance. That is why at Summum we consider it a core part of the project, not an optional add-on: the deployment consultancy always includes the design of the corpus update strategy.