Revamping context-oriented retrieval with hybrid search in Redis 8.4

Feeding the agent: Building context-aware intelligence

Intelligent agents live or die by the quality of their context. The challenge isn’t just finding data, but understanding it: distinguishing what’s relevant now, what forms long-term memory, and what underpins reasoning. Agents must combine symbolic precision (“what was said”) with semantic understanding (“what it meant”) — and do so in real time, without latency or complex post-processing.

Redis has long powered real-time decisioning and contextual retrieval. Its hybrid search already lets developers combine text, tag, and numeric filters to focus vector similarity on what truly matters.

Now, the Redis Query Engine introduces `FT.HYBRID`, a unified in-engine API that fuses full-text and vector results through score fusion, returning one ranked list that balances meaning and match instantly and at scale.

No trade-offs. No post-processing. Just fast, context-aware intelligence, built in Redis.

Why hybrid retrieval matters

Recent research reinforces why hybrid retrieval matters for RAG pipelines and intelligent agents. Studies from Anthropic (2025) and Apple ML Research (2024) show that context retrieval quality alone can shift accuracy by more than 10 percentage points, while hybrid retrieval (text + vector) reduces context failure rates by up to 49% compared to single-mode retrieval.

In hybrid RAG systems such as Blended RAG (2024) and HyPA-RAG (2024), integrating keyword and vector searches improved retrieval recall by 3–3.5× and raised end-to-end answer accuracy by 11–15% on complex reasoning tasks.

These results confirm what developers already observe in production: when retrieval joins lexical precision with semantic understanding, large language models deliver more faithful, grounded, and explainable responses — all without slowing down the query path.

Hybrid search: In action

Redis Query Engine has long supported hybrid-style retrieval via a hybrid policy that constrains the candidate set for vector search using the text/metadata selectors from the same query. In practice, the policy determines how candidates are gathered (e.g., intersecting posting lists with vector neighborhoods or batching strategies) before the fusion step. This keeps latency low while improving precision on large corpora.

The new FT.HYBRID API takes this concept a step further.

Instead of relying solely on pre-filters or post-processing with aggregation pipelines, Redis 8.4 introduces native score fusion combining full-text relevance and vector similarity within a single command execution plan. As simple as:

This means you can now retrieve, filter, and rank results across modalities in one cohesive operation, with consistent normalization and configurable scoring strategies like Reciprocal Rank Fusion (RRF) and Linear Combination.

The foundation for hybrid retrieval was already in place. This release unifies it, simplifies it, and supercharges it.

Context-oriented retrieval with the new FT.HYBRID

When building AI agents, not all memories are created equal. Some facts age quickly, others depend on geography, and many live in the gray space between exact keywords and semantic meaning. Redis Query Engine’s new FT.HYBRID command was designed to handle this full spectrum, allowing you to express contextual intent directly in the query, use and re-use the same structure in your agent framework.

Prioritizing recent memories

When an agent retrieves prior interactions or documents, recency often outweighs raw similarity. With FT.HYBRID, you can easily emphasize freshness by combining time filters with vector relevance surfacing semantically similar but recent events first.

This query fuses BM25 text relevance with vector similarity, then sorts by timestamp giving the agent recent and semantically relevant memories first.

Regionalizing context with GEO and GEOSHAPE

Decision-making agents often operate within a spatial context. Think of a logistics optimizer, travel assistant, or local recommender. FT.HYBRID allows you to filter or rank by geographic proximity while still considering semantic meaning and metadata.

The agent now retrieves “nearby” contexts ranking Seattle cafés by both semantic similarity (“wifi outdoor”) and proximity, using reciprocal rank fusion for precision.

Mixing semantic, fuzzy, and exact matches

Sometimes, you want an agent to understand that “AI workshop” and “machine learning session” are similar, but also to prioritize exact keyword hits when available. FT.HYBRID makes this natural with optional (~) and fuzzy (%) operators combined with vector matching.

Here, Redis fuses fuzzy keyword logic with semantic closeness producing rich, nuanced retrieval for dynamic or user-generated text.

Why this matters for agents

These examples show how FT.HYBRID evolves Redis from a fast search engine into a context engine that can reason over time, place, and meaning simultaneously.

By bringing full-text, vector, and metadata search under a single API with tunable scoring fusion, Redis gives developers the language to express how agents should think, not just what to match.

Context is no longer assembled; it is retrieved. The quality of your agent’s reasoning now starts with the quality of its hybrid search.

Try it yourself

The new Hybrid search API is available starting in Redis 8.4. Locally with Redis.

Once you have Redis running, create an index and start building your own hybrid retrieval workflows.