Pinecone Adds BM25 Full-Text Search to Vector Database for More Reliable AI Agent Retrieval
The company announced the feature on September 9, 2026, following months of private-preview validation. Full-text search adds traditional keyword-based retrieval to the same Pinecone environment already used for vector search.
The result is designed to give AI agents access to both semantic understanding and exact text matching without requiring developers to operate a separate search system.
That could make a significant difference for retrieval-augmented generation, enterprise search and AI agents working with structured business information.
What Is Pinecone Full-Text Search?
Pinecone Full-Text Search is a keyword-search capability built directly into Pinecone Database.
It uses BM25 ranking to find documents based on exact words and terms.
Pinecone also supports text-match filters and Lucene query syntax for more advanced keyword searches. The company says full-text search can operate alongside dense and sparse vector representations inside the same index.
This gives developers another retrieval method without necessarily having to create a separate search infrastructure.
A simplified AI retrieval system can now combine:
Semantic search + keyword search + metadata filtering
That combination is particularly useful when an AI application needs to understand both the meaning of a request and the exact information contained in a document.
Why Vector Search Alone Can Miss Important Information
Vector search has become one of the foundations of modern AI applications.
Text is converted into embeddings, and those embeddings allow systems to retrieve information based on semantic similarity.
For example, a user might search:
"Why did the payment fail?"
A vector database can retrieve documents discussing failed payments even if they do not contain exactly those words.
That is extremely useful.
But semantic similarity can become a problem when the user needs an exact match.
Imagine a customer support agent searching for:
PROD-001
A semantic search system could potentially return documentation for similar product numbers.
That may be unacceptable if the customer is asking about one specific product.
The same problem can occur with:
- Error codes
- Product SKUs
- Customer IDs
- Legal citations
- Version numbers
- Names
- Serial numbers
- Technical identifiers
- Database fields
- API endpoints
These cases often require exact or near-exact keyword retrieval.
Pinecone's new capability is designed for that situation.
BM25 Brings Exact Keyword Ranking Into Pinecone
The foundation of Pinecone's full-text search is BM25, a widely used ranking approach for keyword search.
Unlike embedding-based retrieval, BM25 does not need to understand the semantic meaning of the entire query.
Instead, it evaluates how relevant the query's terms are to the text being searched.
This makes it particularly useful when specific words matter.
Pinecone says full-text search supports BM25 ranking across text fields and can be combined with filters and other retrieval signals.
For developers, that means an AI application can use the retrieval method that best matches the question instead of forcing every query through embeddings.
One Database Can Now Handle Different Retrieval Signals
One of the most important aspects of the announcement is that Pinecone is not presenting full-text search as a completely separate product.
The company says full-text search can exist in the same index as dense vectors and sparse vectors.
That means an application can maintain different kinds of retrieval signals within the same database architecture.
For example:
Dense vectors can help with semantic similarity.
Full-text search can handle exact keywords and phrases.
Sparse vectors can provide another lexical retrieval method.
Metadata filters can restrict results based on structured information.
This gives developers more flexibility when building AI retrieval systems.
Why AI Agents Need Both Meaning and Exact Text
AI agents increasingly work with large amounts of enterprise information.
A customer-support agent may need to find a specific order.
A coding agent may need to locate an exact error message.
A financial assistant may need to retrieve a specific transaction identifier.
A healthcare application may need to find a particular medical term.
A developer documentation assistant may need to identify an exact API name.
In each case, semantic search can be useful, but exact text can also be critical.
Pinecone's argument is that AI retrieval should not have to choose one approach for every query.
Instead, the application can use the retrieval signal that fits the information being searched.
Pinecone Supports Lucene Query Syntax
The new system also supports Lucene query syntax.
That gives developers more control over keyword searches than a simple text query.
Lucene syntax can be used for operations such as Boolean queries, phrase matching and other advanced search conditions.
Pinecone's documentation says full-text search supports query-string searching through Lucene syntax, alongside BM25 text ranking.
This is useful for applications that need more precise search behavior.
For example, an enterprise search system could require several terms to appear together, search for an exact phrase or apply more structured query conditions.
Text Match Filters Add Another Layer of Precision
Pinecone also supports text-match filters.
The documentation lists operators including:
$match_phrase$match_all$match_any
These allow applications to apply more precise text conditions while searching.
That matters for AI systems because retrieval quality often determines the quality of the final answer.
If the retrieval layer provides the wrong information, even a powerful language model can produce a poor response.
Improving retrieval precision can therefore improve the entire AI pipeline.
Full-Text Search Works Alongside Semantic Search
Pinecone is not suggesting that full-text search should replace vector search.
The two approaches solve different problems.
Consider a technical support application.
A user might ask:
"How do I fix the database connection timeout?"
Semantic retrieval could identify relevant documentation even if the exact wording is different.
But if the user instead asks:
"What does error DB-504 mean?"
Exact keyword retrieval may be much more appropriate.
The ideal AI system can support both.
That is the direction Pinecone is taking.
Hybrid Retrieval Becomes More Practical
Combining different search approaches can create a more robust retrieval pipeline.
An AI application could retrieve documents using semantic similarity and then use keyword matching or text filters to narrow the results.
For example, a support system might first identify documentation related to a customer's problem and then prioritize documents containing the exact product code mentioned in the question.
This can reduce the risk of retrieving something that is conceptually similar but operationally incorrect.
Pinecone's documentation describes full-text, semantic, sparse and hybrid search as different retrieval approaches that can be selected according to the application's needs.
Pinecone Says Developers No Longer Need a Separate Search Cluster
Before integrated full-text search, developers needing exact keyword retrieval could operate a dedicated search engine alongside their vector database.
That can create additional infrastructure requirements.
A separate search system may require:
- Capacity planning
- Infrastructure management
- Scaling
- Memory management
- Sharding
- Index maintenance
- Version upgrades
- Monitoring
Pinecone's pitch is that developers can avoid some of that additional operational complexity by putting full-text search directly into Pinecone Database.
For teams already using Pinecone for vector retrieval, that could simplify their architecture.
The Documents API Is Central to the New Capability
Pinecone's full-text search works through its document-oriented data model.
Developers can store JSON documents and define text fields for full-text indexing.
The documentation says a schema can contain strings configured for full-text search alongside dense or sparse vector fields.
That means developers can structure information according to the different retrieval methods they need.
A single document might contain:
- Title
- Description
- Product ID
- Technical documentation
- Embedding
- Metadata
Different parts of that information can then support different retrieval strategies.
Full-Text Search Is Especially Useful for Enterprise AI
Enterprise AI applications often deal with information that contains many exact identifiers.
Business systems are full of:
- Invoice numbers
- Contract IDs
- Customer numbers
- Product codes
- Employee IDs
- Database fields
- Error codes
- Internal project names
Semantic similarity is valuable for understanding natural-language requests.
But exact matching remains essential for structured business information.
This makes full-text search a particularly relevant addition for enterprise AI applications.
Better Retrieval Could Improve RAG Systems
Retrieval-augmented generation, commonly known as RAG, depends heavily on finding the right information before an AI model generates its response.
The process usually looks something like:
User question → Retrieval → Relevant documents → AI model → Answer
If retrieval fails, the language model may receive incomplete or incorrect context.
Adding multiple retrieval methods can potentially make this pipeline more robust.
Full-text search can handle exact terminology.
Vector search can handle semantic meaning.
Metadata filters can restrict the search space.
The AI model can then generate an answer using the retrieved context.
Pinecone is positioning its database as infrastructure for this type of knowledgeable AI system.
Pinecone Is Moving Beyond Being Just a Vector Database
The full-text search launch also reflects a broader evolution in Pinecone's product strategy.
The company increasingly describes Pinecone as a knowledge platform for AI agents, rather than simply a vector database.
Its website emphasizes giving agents access to knowledge through retrieval infrastructure.
That positioning is important as AI agents become more dependent on external information.
An agent can have strong reasoning capabilities but still require access to company documentation, customer information, product data and other knowledge sources.
The retrieval layer becomes part of the agent's overall intelligence stack.
How Pinecone Full-Text Search Differs From Traditional Search
Traditional search engines remain highly effective at keyword retrieval.
The difference here is integration.
Pinecone is bringing exact keyword retrieval into an environment already designed around vector and AI retrieval.
That means developers can build applications where different search signals work together rather than maintaining completely separate systems.
The objective is not to replace traditional search engines in every situation.
Instead, Pinecone is targeting AI applications that need both semantic understanding and exact retrieval.
What Developers Should Consider
Pinecone Full-Text Search could be particularly useful when an application deals with exact terms.
Developers should consider it for applications involving:
- Product catalogs
- Customer-support systems
- Technical documentation
- Code search
- Enterprise knowledge bases
- Legal documents
- Financial records
- Inventory systems
- Error-code databases
- AI agents using structured business data
However, full-text search is not automatically the best option for every query.
If the goal is understanding the meaning of a question and finding conceptually related information, dense-vector search can remain valuable.
The strength of Pinecone's approach is the ability to use both.
What This Means for AI Tools
The AI-tools market is increasingly moving away from simple model-centric architectures.
Developers are now building systems that combine:
Models + Agents + Tools + Retrieval + Data
The retrieval layer is becoming increasingly important.
A capable AI agent without reliable access to information can still make mistakes.
Pinecone's latest release focuses on one specific part of that problem: helping AI systems retrieve the right information using the right search method.
That may not be as visible as a new frontier AI model, but it is an important infrastructure development.
Pinecone's decision to make full-text search generally available is significant because it addresses a practical weakness in AI retrieval.
Vector search is excellent for understanding meaning.
Keyword search remains essential when exact terms matter.
By putting BM25 full-text search, text-match filters and vector-based retrieval into the same database environment, Pinecone is giving developers more options for building reliable AI search and agent systems.
The biggest benefit may be architectural simplicity.
Teams can potentially avoid maintaining separate infrastructure just to handle exact keyword searches while still benefiting from semantic retrieval.
For AI agents, RAG applications and enterprise knowledge systems, that combination could become increasingly important.
As AI applications move from answering general questions toward performing real work, retrieval accuracy becomes just as important as model intelligence.
Pinecone's latest update is a clear example of that shift.
FAQs
What is Pinecone Full-Text Search?
Pinecone Full-Text Search is a generally available keyword-search capability in Pinecone Database that uses BM25 ranking and supports text-match filters and Lucene query syntax.
Is Pinecone Full-Text Search different from vector search?
Yes. Full-text search focuses on keyword and phrase matching, while vector search retrieves information based on semantic similarity. Pinecone supports both approaches.
