Your AI search pipeline is broken. This open-source framework fixes the plumbing.

Building search infrastructure for AI applications has become a category of engineering pain that most teams accept as unavoidable. The pattern is familiar: choose an ingestion tool, pick a vector database, wire up a retrieval library, and bolt on evaluation scripts that don't share assumptions with any of the above. Each component speaks its own data language, expects its own document format, and breaks when the pipeline upstream changes.

Today's public preview release of Search Toolkit directly addresses this fragmentation. The open-source framework, built by a team that has developed enterprise search infrastructure for years, brings ingestion, retrieval, and evaluation under a single composable interface. It runs wherever a team's infrastructure runs, cloud, on-premises, or edge, and carries the explicit pitch that teams should spend their engineering time on search quality, not pipeline assembly.

"We built it because teams building search infrastructure still spend too much engineering time on plumbing," the release team wrote. "Most stitch together separate tools for ingestion, retrieval, and evaluation, each with its own interface and its own assumptions about data."

The real cost of integration churn

The underlying problem is not that the individual tools are bad. It's that each integration point between them creates a maintenance surface that grows with every new data source, every schema change, and every evaluation run. Teams report spending weeks just to reach the point where they can run a single query against their own data. Measuring retrieval quality often requires yet another toolchain, with its own data shape assumptions that may or may not match what the ingestion pipeline produced.

Search Toolkit attacks this at the architecture level. All modules, ingestion, retrieval, evaluation, share a common configuration interface. Swap a retriever, add an evaluator, change the chunking strategy. The rest of the pipeline adapts without manual rewiring.

Three problems, one framework

The release breaks down the use cases into three categories, each representing a distinct pain point in current enterprise search deployments.

Enterprise search fragmentation. Most organizations don't have a single search problem. They have a dozen: internal wikis, support ticket systems, document repositories, file storage, codebases. Each source has different structure and metadata, requiring different processing to index well. The standard result is either isolated indexes that can't be searched together or a brittle custom layer that becomes its own maintenance burden. Search Toolkit provides consistent processing and indexing patterns across source types, so adding a new source doesn't mean rebuilding the pipeline.

RAG quality isolation. When a retrieval-augmented generation system returns poor results, teams typically lack a clean way to determine whether the problem is retrieval or generation. The reflex response is to tweak prompts, adjust chunking, and swap models, all without knowing whether the retriever is surfacing the right context. Search Toolkit includes built-in evaluation metrics (recall, precision, MRR, NDCG) that measure retriever performance independently, allowing teams to isolate retrieval quality from generation quality.

Domain-specific retrieval. Off-the-shelf retrievers are trained on general-purpose text and struggle with the specialized terminology and document structures of legal filings, medical records, codebases, and financial disclosures. Teams building domain-tuned retrieval often end up constructing custom infrastructure from scratch. Search Toolkit's configurable retrievers, BM25 sparse, dense embedding-based, and hybrid configurations, allow teams to tune for their particular domain without rebuilding the retrieval layer.

Search in an agentic architecture

The framework arrives at a moment when the agent paradigm is reshaping how enterprises think about retrieval. Agents working on enterprise tasks need access to enterprise context, and they make retrieval decisions autonomously at high volume. The quality of the infrastructure directly affects every downstream step.

Search Toolkit supports two retrieval paths for agents: indexed semantic search across large document corpora and live data pull from source systems through MCP integrations. An agent can query an indexed corpus when it needs to search across a large body of content, and pull live data from a CRM, code repository, or productivity tool when it needs the latest state.

Battle-tested deployments

The framework has been designed for advanced enterprise use cases and tested across financial services, manufacturing, public sector, and media verticals. One notable deployment involves CMA CGM, which uses Search Toolkit alongside Voxtral to help journalists detect fake news. The pipeline processes audio from three distinct data sources and returns alerts within 15 seconds end to end.

Search Toolkit is released as open source and available now through a starter app template that provides pre-configured Vespa indexing, hybrid retrieval, and sample data. The full documentation covers schema management, relevance optimization, and advanced retrieval features including LLM query rewriting and reranking.