November 2022 Summaries

Bi-encoders are an efficient architecture for semantic Question Answering (QA) systems, where both questions and answers are embedded into a vector space, allowing semantically similar pairs to be close to each other. Using Cohere's co.embed API and Qdrant, a vector search database, simplifies the setup of such systems, offering scalable and maintainable solutions without the need for personal infrastructure. This approach is exemplified by implementing a QA system on biomedical data using the pubmed_qa dataset, where embeddings of questions and answers are stored in Qdrant and queried to find relevant answers. The system's performance is evaluated using top-k accuracy metrics, showing satisfactory results, with accuracy increasing as more results are considered. The combination of Cohere and Qdrant provides a convenient SaaS solution for building and querying QA systems, allowing for customization and fine-tuning with specific datasets for improved domain-specific performance.

Qdrant, an efficient vector database, initially supported only keyword filters for semantic searches, but since version 0.10, it has enabled full-text filtering capabilities, allowing users to apply more complex constraints in conjunction with other filter types. Full-text filters can be used without an index, performing substring matches on individual query terms, or with an index, offering more options like choosing a tokenizer and setting parameters such as token length and case sensitivity. The primary advantage of using full-text indexes is improved query performance, as demonstrated in a benchmark using the H&M dataset, where indexed fields offered substantial performance gains when queried frequently.