Qdrant

Keyword indexes in Qdrant are built for exact matching. They’re ideal for values like URLs, file paths, SKUs, and unique identifiers. The challenge? If you wanted to find values that start with something, for example, all URLs beginning with “https://qdrant”., you had two options: - Scan every value in the payload, or - Switch to a text index, which tokenizes the value and no longer supports exact matching. Qdrant 1.19 solves this. You can now enable prefix matching on keyword indexes with "prefix": true in the index configuration. After that, prefix filters are served directly from the index, giving you fast prefix lookups while keeping exact matching. Full release here: https://lnkd.in/g_za6_Ca

We just added a new item to the Qdrant product list: fully managed Qdrant Cloud, Qdrant Hybrid Cloud/BYOC, Qdrant Enterprise (on-prem) deployment, and Qdrant Edge. A waiting list for Qdrant.to/Serverless 🥷

BM25 scoring relies on IDF (Inverse Document Frequency). In simple terms, the rarer a word is in your data, the more important it becomes during search. In a multi-tenant collection, IDF was previously calculated across all tenants combined. That meant if Tenant A and Tenant B used very different vocabularies, their statistics were mixed together. As a result, a word that was rare for one tenant might not be treated as rare, leading to less accurate BM25 scores. Qdrant 1.19 fixes this. You can now calculate IDF statistics per tenant instead of across the entire collection. Each tenant’s BM25 scores are now based only on their own data, resulting in more accurate lexical search without requiring separate collections. Full release here: https://lnkd.in/g_za6_Ca

Native TurboQuant datatype Turbo4 is definitely a highlight of the Qdrant v1.19 release. However, a few more things were introduced along with it that are worth mentioning. ⁃ Memory tiers: Every collection component (vectors, HNSW graph, quantized vectors, sparse index, payloads) used to have its own on-disk/in-RAM knob, with inconsistent names across the board. Now there's one memory parameter with three tiers: pinned (always in RAM), cached (on disk, pre-warmed into OS cache at startup), and cold (lazy load on first access). Same config shape everywhere. HNSW links can now be pinned, and quantized vectors get their own independent tier from the originals. ⁃ Per-tenant IDF: If you're running BM25 in a multi-tenant setup, term rarity used to get computed across the whole shard. Tenant A's vocabulary was polluting tenant B's scores. You can now scope the IDF corpus down to a single tenant. ⁃ Keyword Filtering: Two additions: prefix matching on keyword fields (finally, "starts with" queries on exact-match indexes without a full scan), and a new slice condition that splits a collection into deterministic, non-overlapping subsets. Good for parallel scrolling across workers or pinning down a reproducible sample for benchmarking. ⁃ Web UI: The Collection Visualizer got rebuilt server-side with a WebAssembly layout engine and WebGL2 rendering, so it now handles tens of thousands of points instead of a few thousand. You can also click a point for its nearest neighbors or shift-drag to select a region. Resharding now shows live progress, and payload indexes are manageable directly from the UI. Full writeup: https://lnkd.in/gjTeCsT8

Qdrant 1.18 introduced TurboQuant, a quantization technique that speeds up vector search by keeping a 4-bit compressed copy of your vectors in memory while storing the original float32 vectors on disk for rescoring. Great recall, but two copies of every vector. We just released Qdrant 1.19, which introduces Turbo4, a new datatype that stores only the 4-bit representation, removing the original float32 copy entirely. That reduces storage from 36 bits per coordinate to just 4 bits, a 9× storage reduction. With less data to read and write, search throughput also improves. The tradeoff is simple: without the original vectors, rescoring isn’t possible. If storage efficiency is your priority, Turbo4 is the right choice. If you need the highest possible recall, TurboQuant with full-precision storage is still the better option. The same storage savings apply to multi-vector collections, making ColBERT-style late interaction search significantly more space-efficient. Full release here: https://lnkd.in/g_za6_Ca

Qdrant's WebUI 𝗶𝘀 𝘁𝗵𝗲 𝗳𝗶𝗿𝘀𝘁 𝘁𝗵𝗶𝗻𝗴 𝗜 𝗿𝗲𝗰𝗼𝗺𝗺𝗲𝗻𝗱 𝘁𝗼 𝗽𝗲𝗼𝗽𝗹𝗲 𝗹𝗲𝗮𝗿𝗻𝗶𝗻𝗴 𝘃𝗲𝗰𝘁𝗼𝗿 𝘀𝗲𝗮𝗿𝗰𝗵. If you are able to figure this out, the rest won't be so hard. The visualization feature lets you debug your embedding model and the entire extraction and vectorization pipeline. If you click on a point, your vector's ANN matches will stand out, so that you can see if the data is properly embedded and indexed. 𝗔𝘀 𝗼𝗳 𝘃𝗲𝗿𝘀𝗶𝗼𝗻 𝟭.𝟭𝟵, 𝗪𝗲𝗯𝗨𝗜 𝗻𝗼𝘄 𝗻𝗼𝘄 𝘀𝘂𝗽𝗽𝗼𝗿𝘁𝘀 𝗹𝗮𝗿𝗴𝗲 𝗱𝗮𝘁𝗮𝘀𝗲𝘁𝘀. (https://t.ly/QGDrU) Both Neil Kanungo and I tried it last night and were able to load 20K+ vectors in a few seconds. Not sure why but this little dashboard always gets me fired up!

Can your Slack remember what your team already knows? Join Qdrant and cognee in Berlin for an evening of building AI memory that goes beyond context windows. At this hands-on hack night, you’ll build a memory layer that turns your Slack workspace into something agents can actually reason over, using cognee for structured memory and Qdrant for fast semantic retrieval. - €1,200 prize pool: - Plus, you’ll get swag, API keys, sample datasets, and support from the cognee and Qdrant teams throughout the event. Date: August 14 Location: Paul-Lincke-Ufer 39-40/hof 4, Berlin Register now: https://lnkd.in/dmeCRCgD #Qdrant #Cognee #AI #AgenticAI #RAG #VectorSearch #BerlinTech #HackNight

Qdrant 1.19 is here! Below are the main updates. - TurboQuant Datatype Version 1.18 shipped TurboQuant as a quantization layer. 1.19 enables the new Turbo4 datatype, which keeps only the 4-bit representation. That takes 36 bits per coordinate (the float32 original plus the 4-bit copy) down to four, a ninefold storage reduction, and fewer reads and writes per operation lifts throughput. - Memory Tiers One memory parameter now replaces on_disk, always_ram, and on_disk_payload, and it works the same way on every component of a collection. - Per-Tenant IDF Statistics You can now narrow the corpus to a single tenant, so term rarity reflects that tenant's data instead of the whole dataset. - Filtering Prefix matching on keyword indexes means no more full payload scans, and no more trading exact matching for a text index. One upgrade note: the legacy /search, /recommend, and /discover endpoints are gone. Migrate to /query before you upgrade. The Turbo4 datatype trades rescoring capability for storage savings, without a full-precision copy, top candidates can't be rescored. Use it when disk reduction is the priority. Full release blog here: https://lnkd.in/g_za6_Ca

Why does your RAG agent get worse over time? Join Dylan Couzon and Rishabh from Future AGI tomorrow as they build, debug, and optimize a RAG agent live, covering embedding migrations, deduplication, reranking, and evaluation to measure what actually improves retrieval quality. Time: 6:30 PM CEST | 9:30 AM PDT | 10:00 PM IST Register now: https://lnkd.in/gfKUauMH

We just shipped Turbo4 in Qdrant 1.19 It's a new data type for primary vector storage of dense vectors: 4 bits per dimension instead of the usual 32. That's 8x less memory, before you even touch quantization. The trick isn't just "store fewer bits." Naive 4-bit rounding wrecks recall because some dimensions carry way more signal than others, and they lose it disproportionately when you truncate. Turbo4 fixes this by first applying a random rotation: every vector is mathematically rotated, so information spreads evenly across all dimensions. Then each rotated value gets bucketed into one of 16 levels (all 4 bits can represent). No dimension gets unfairly punished. This is TurboQuant, the vector quantization method Google Research published earlier this year. We already used it for quantization in 1.18. Turbo4 takes the same rotation-and-bucket idea and turns it into a storage format you can set directly on a collection to spare disk space on original vectors. Docs: https://lnkd.in/d6a22R53