Published at: 2026-09-17
Semantic retrieval index overview
RAG and the knowledge index
A knowledge index is like a smart catalog for a huge enterprise repository. It distills massive documents and data into a structured, easy-to-search address book. AI uses it to locate relevant knowledge quickly.
The index and RAG work together. RAG (Retrieval-Augmented Generation) lets a large model answer questions. When RAG asks a question, the knowledge index is its instant data source. RAG queries the index, finds the most relevant fragments, and generates an accurate answer based on them.
Why build an index, and what does it mean for your enterprise?
- Efficiency and cost: Enterprise knowledge bases can reach terabytes. Feeding them directly to a large model is slow and expensive. An index pre-organizes information for millisecond retrieval.
- Accuracy: Answers stay grounded in the latest internal knowledge (product manuals, customer cases), which prevents hallucination.
- Expertise: A general-purpose large model becomes an enterprise-specific expert and unlocks high-quality internal knowledge Q&A.
Semantic segmentation strategies
ShareCRM uses the following strategies in semantic segmentation to improve recall accuracy.
Select a title field for knowledge
When segmenting an object, select a title field. During retrieval, the system searches the title field in addition to semantic and full-text recall. It treats the segments that match the title field as a separate recall path, then reranks all paths to produce the most complete answer.
Tip: Choose a field with business meaning that users query most often.
Add a chunk field
Segmentation can include a chunk field. The chunk field value is added to each segment as semantic reinforcement.
Tip: Choose a field with business meaning that users query most often.
Document hierarchy in segmentation
For document-type indexes, the system adds the document hierarchy of each segment to every chunk (for documents with a hierarchy). Hierarchy adds relationships and topic context to segments, which improves recall accuracy and quality. If a question matches a hierarchy node, all segments under that node are recalled to generate the answer, which improves recall and precision.