Why Antirez Is Returning to Redis: Insights on Licensing, AI, and Vector Search

Why Antirez Is Coming Back

On December 10, Redis Labs CEO Rowan Trollope wrote on the Redis blog that Antirez (Salvatore Sanfilippo), the original creator of Redis, has re‑joined the company and will serve as a Redis ambassador.

In his accompanying article “From where I left”, Antirez explains why he decided to return after a four‑year break, what he hopes to achieve, and how his perspective on open‑source licensing has evolved.

Reasons for Returning

He notes that he stopped looking at Redis source code for about 1,620 days, only compiling it when needed, and that his break was motivated by a desire to spend more time with family, write, and explore other technologies such as embedded projects, neural networks, and Telegram bots.

He also observes a split in the Redis community and wonders whether he can act as a bridge between the company and its users, helping to refocus development on the core.

Licensing Discussion

Antirez clarifies that his return was initiated by him, not by the company, and that he is not receiving a huge salary—just a normal compensation package, while still holding a modest amount of Redis stock options.

He expresses limited objection to the recent license change, stating that the community split is not solely caused by licensing. The new SSPL‑like license still permits most uses of Redis, provided the software is not offered as a service without also open‑sourcing the surrounding orchestration layer.

He argues that open‑source is not the only way to develop software and that the license change aims to prevent cloud providers from simply repackaging Redis as a service without contributing back.

AI, Large Models, and Vector Indexing

Redis is actively developing vector capabilities to support AI workloads. Antirez, who built his first neural‑network library in 2003, now uses tools like Claude AI to accelerate development, write code, and benchmark 8‑bit quantized vector dot‑product performance.

He believes that Retrieval‑Augmented Generation (RAG) is not the only valuable AI application and that embedding‑based vector search will remain important. Redis’s approach lets developers decide trade‑offs by exposing vector indexes as ordered‑set‑like structures.

Prototype Vector Commands

He is prototyping new commands (VSIM, VADD, VCARD) that treat the score of an ordered set as a multi‑dimensional vector and perform K‑NN matching. The proof‑of‑concept code is shown below.

VSIM top_1000_movies_imdb ELE "The Matrix"  WITHSCORES
 1) "The Matrix"
 2) "0.9999999403953552"
 3) "Ex Machina"
 4) "0.8680362105369568"
 5) "Akira"
 6) "0.8635441958904266"
 7) "District 9"
 8) "0.8631418347358704"
 9) "The Martian"
10) "0.8608670234680176"
11) "The Bourne Ultimatum"
12) "0.8599717319011688"
13) "The Bourne Supremacy"
14) "0.8591427505016327"
15) "Blade Runner"
16) "0.8585404753684998"
17) "Metropolis"
18) "0.8572960793972015"
19) "Inception"
20) "0.8521313071250916"

These commands are built on ordered‑set concepts but introduce multi‑dimensional scores (embeddings) and K‑NN matching, with plans to add a STORE option to save results as ordered sets.

Looking Forward

Antirez sees Redis as a “Lego set for programmers” rather than a finished product, emphasizing that new features must fit both use‑case and internal design. He hopes his return will help steer Redis back toward core development, improve licensing clarity, and advance AI‑driven vector search capabilities.

He concludes with a simple statement: “I’m back.”