Agentic Retrieval Systems

5/19/2026

This post details GoPerfect's implementation of an agentic recruiting workforce, highlighting the use of Qdrant Cloud for advanced retrieval. Key technical contributions include: 1. Moving from purely semantic dense embeddings to a multi-vector representation in Qdrant to support more granular retrieval and disambiguate nuanced job roles. 2. Implementing hybrid search (semantic and keyword) to combine different signals for improved accuracy. 3. Orchestrating LLMs on top of Qdrant's retrieval results to form an agent loop for complex decision-making and ranking. 4. Achieving near-perfect match accuracy (99.993% in benchmarks) and sub-10-second complex search completion times through this architecture. 5. Describing the production pipeline involving data ingestion into Qdrant with structured multi-vector points, retrieval, and LLM orchestration for agentic workflows.

3/31/2026

This post introduces 'Qdrant skills,' a new capability that encodes solutions architect knowledge for AI agents. These skills provide diagnostic decision trees for agents to navigate Qdrant's complex feature set, addressing issues like memory optimization, search quality, and scaling. The skills are organized by problem rather than feature, offering guidance on 'when' and 'why' to use specific Qdrant functionalities, and include 'what NOT to do' advice. This complements existing agentic retrieval system capabilities by providing the intelligence layer for agents to effectively manage and optimize Qdrant deployments.

3/24/2026

Introduces a free, advanced course on Multi-Vector Search, detailing its structure, learning objectives (late interaction, ColBERT, ColPali, quantization, MUVERA indexing), and hands-on implementation using Qdrant. The course aims to bridge the gap between theoretical understanding of multi-vector retrieval and practical production-level implementation.

2/3/2026

This post details the winners of the Sketch & Search hackathon, showcasing how Qdrant was used as a memory layer in agentic retrieval systems. The winning projects (Prometheus, Roast My Snack, AutoScape) utilized Qdrant for storing embeddings of past assets, prompts, and outcomes to enable reuse, consistency, and grounded outputs. Prometheus used Qdrant to search prior targets for prompt/camera templates. Roast My Snack used Qdrant for semantic search to inform age-adaptive roasts. AutoScape employed a Qdrant-powered RAG system with a curated plant database for grounded landscape designs.

1/28/2026

This post introduces two complementary approaches for enabling AI agents to interact with APIs: SKILL.md for providing explicit knowledge about API usage (preventing known mistakes like using deprecated methods or incorrect configurations) and a REPL-first approach for dynamic discovery of API capabilities and schema (addressing unknown unknowns specific to the user's environment). The post details how these approaches can be combined to create a more robust agent workflow, where agents first consult SKILL.md for guidance and then use a REPL to explore the actual API state before generating code. It also provides a minimal implementation example for a REPL tool and discusses the broader pattern applicable to any API with deprecated methods or user-specific state.

1/28/2026

This post details how Anima Health uses Qdrant as a core infrastructure component for vector search, similarity analysis, and agentic AI workflows in their clinical document intelligence platform. Specifically, it highlights the use of Qdrant for retrieving SNOMED codes by representing them as vector embeddings with enriched metadata, and how this retrieval layer within an agentic pipeline narrows the search space for high-confidence recommendations. The post also emphasizes Qdrant's role in identifying patterns across documents while preserving patient privacy through embeddings, and discusses the benefits of Qdrant's deployment flexibility (self-hosted, region-controlled), cost predictability, payload-based filtering, and multivector support for healthcare applications. Future directions for Anima Health include expanding agentic systems, modeling longer-term patient state, and exploring temporal relevance.

12/17/2025

This post details Qdrant's 2025 recap, focusing on its evolution to power the agentic era. Key technical advancements include: Advanced Retrieval with score-boosting reranking, full-text filtering (multilingual tokenization, stemming, phrase matching), ACORN algorithm for filtered HNSW queries, Maximal Marginal Relevance (MMR) for diversity, and ASCII folding. Performance & Resource Optimization features include GPU-accelerated HNSW indexing, inline storage for quantized vectors, a custom storage engine for low-latency access, incremental HNSW indexing for upsert-heavy workloads, HNSW graph compression, and expanded quantization options (1.5-bit, 2-bit, asymmetric). Enterprise Scaling & Isolation enhancements cover tiered multitenancy, SSO/RBAC, granular database API keys, Terraform-enabled Cloud API, and conditional updates. Deployment Flexibility is addressed with Qdrant Cloud Inference for unified embedding generation and search, and Qdrant Edge for on-device retrieval.

12/11/2025

This post announces a new DeepLearning.AI course focused on multi-vector image retrieval using ColPali and MUVERA, building upon previous work in retrieval optimization. It details the course content, which includes understanding multi-vector embeddings, using ColBERT for text retrieval, mastering ColPali for image patch-level representations, optimization techniques like quantization and pooling, MUVERA's approach to efficient representations, and building multi-modal RAG pipelines with Qdrant's HNSW search.

10/10/2025

This post details TrustGraph's technical implementation of enterprise-grade agentic AI using Qdrant. It highlights their architecture which combines Apache Pulsar for resilient streaming, RDF for graph-native knowledge modeling, and Qdrant for vector search. Key technical aspects include: extracting facts and embedding entities for dual representation in Qdrant and the knowledge graph; a query process that embeds queries, retrieves nearest entities from Qdrant, expands to a subgraph, and feeds this curated context to an LLM; and an agentic AI framework that orchestrates GraphRAG, template-driven queries, and tool invocation. The post emphasizes production-grade outcomes like determinism, resilience, scalability, and data sovereignty achieved through this integration.

9/4/2025

Fieldy AI migrated from Weaviate to Qdrant for their AI memory platform due to reliability issues with Weaviate (5xx errors). The migration involved reusing the existing vector schema, batching embeddings locally before ingestion, disabling indexing during bulk import, and updating API calls to use Qdrant's gRPC interface. They implemented hybrid search with BM25 and dense vector retrieval via Reciprocal Rank Fusion (RRF). The migration resulted in 100% reliable recall, reduced latency, virtually eliminated network errors, and a two-thirds reduction in infrastructure costs while scaling to tens of millions of embeddings.

7/31/2025

This post details how PortfolioMind implemented real-time crypto intelligence using Qdrant. It describes their challenge of moving beyond static insights to model dynamic user curiosity by transforming user interactions into multivector user-intent models. The system ingests diverse data (news, tokenomics, whale behaviors, etc.), embeds it with rich metadata, and uses HDBSCAN clustering to identify user-specific micro-interests. Qdrant was chosen for its fast, filterable searches with metadata, native multivector support, low-latency retrieval, and managed services. The results show a 70% drop in latency, a 58% increase in interaction relevance, and a 22% rise in user retention. Future plans include cross-user curiosity mapping, temporal drift tracking, and improved cold-start onboarding.

7/16/2025

Qdrant is now available on AWS Marketplace within the new AI Agents and Tools category. This simplifies the process for customers to discover, buy, and deploy Qdrant for AI agent development, accelerating agentic workflow development. The listing includes Qdrant Cloud and Qdrant Hybrid Cloud options for private AWS deployments. The post emphasizes how this availability helps reduce agent failure, accelerate agentic RAG pipelines, and enable real-time, context-rich reasoning by providing a streamlined way to access Qdrant's vector search engine.

7/14/2025

Introduces a novel approach to modeling aesthetic taste by representing users not as single points but as collections of evolving taste clusters. This involves converting user-art interactions into signed weights, generating artwork embeddings, clustering positive and negative interactions independently using HDBSCAN, scoring clusters by recency using an exponential decay function, and representing users as multivectors (positive and negative). The system leverages Qdrant's Recommendation API with multivector search to find artists with overlapping preferences while minimizing similarity to rejected content, demonstrating a flexible and universal recommendation pipeline.