Build an intelligent search pipeline for grounded AI | THRSP1066
Andy Tran presents the components of a search pipeline and demonstrates how Cohere’s Parse, Embed, and Re-Rank models support document processing and retrieval. The session is aimed at developers building AI systems that need relevant company information at generation time.
Search pipeline overview
The pipeline starts with a user question or instruction, retrieves relevant information from company datasets, and supplies that information to the model when it generates a response.
The session introduces four components:
- Parsing: Extract information from source documents so it can be used in retrieval.
- Embedding: Represent content as vectors for similarity search.
- Hybrid search: Include hybrid search as part of the retrieval pipeline.
- Re-ranking: Improve the ordering of retrieved results, particularly for complex queries.
Cohere’s Parse, Embed, and Re-Rank models address different stages of this process.
Parsing documents
The Parse demonstration covers optical character recognition (OCR), visual extraction, charts, tables, and structured document parsing. These capabilities help make information contained in documents available to downstream search.
Tran also discusses the benefits of purpose-built parsing models compared with larger, more expensive language models.
Multimodal retrieval with embeddings
The Embed overview introduces unified multimodal vector search for text and images. Demonstrations show how retrieval can support different types of content and queries:
- Retail and e-commerce: Multimodal search for fashion products.
- Manufacturing: Image-based retrieval for searching technical manuals.
These examples extend retrieval beyond text-only document search.
Re-ranking for multi-condition queries
The Re-Rank segment focuses on improving retrieval accuracy and handling complex queries. A Nike product-search demonstration shows how re-ranking improves results for searches that must satisfy multiple conditions.
The distinction is between finding candidate results and ordering those results according to how well they meet the user’s request.
Session chapters
- 00:00: Introduction to intelligent search and retrieval pipelines
- 00:32: Pipeline components: parsing, embedding, hybrid search, and re-ranking
- 01:18: Introduction to Cohere Parse, Embed, and Re-Rank
- 01:40: Parse demonstration: OCR, visual extraction, charts, tables, and structured documents
- 03:34: Purpose-built parsing models compared with larger, more expensive LLMs
- 05:03: Embed overview: unified multimodal vector search for text and images
- 06:12: Retail and e-commerce fashion-search examples
- 07:26: Manufacturing and technical-manual image-based retrieval
- 10:15: Re-Rank: retrieval accuracy and complex-query reasoning
- 11:49: Nike demonstration: improving multi-condition product searches