1
Prepare your documents
As your documents are indexed, they need to be prepared for Lucidworks AI. This includes and .Both steps run in one stage called the LWAI Chunker Stage. Add it to your index pipeline anywhere before the Solr Partial Update Indexer stage.
LWAI Chunker stage in the correct position in the index pipeline.
<doc.embedding_t>.Index some documents and continue to the next step.Choose a chunking strategy
Choose a chunking strategy
Lucidworks AI supports several chunking strategies. Each one splits text differently and is suited to specific content types. The strategy you choose affects both performance and response quality. Lucidworks recommends starting with the Sentence strategy and experimenting from there.
The sentence chunker splits text into fixed-size chunks based on a set number of sentences. This approach is simple and consistent, making it a good choice for most use cases where the structure of the text is clear.Best for:
- FAQs
- Help articles
- Structured documentation
Choose a model for indexing
Choose a model for indexing
The model you choose at indexing time controls how your documents are vectorized. These vectors are what Lucidworks AI uses to compare queries to document chunks. A good model captures the meaning of your content clearly and compactly. Model quality affects how accurately the system retrieves relevant chunks during search or generation. Model size affects indexing speed, storage cost, and system performance.Start with a clear goal for your retrieval use case. If you need fast responses and can tolerate a small drop in relevance, choose a smaller model. If you need precise chunk matching across large or complex content, use a higher-quality model—even if it’s slower.Common starting points
- e5-base-v2: Balanced quality and speed. Works well for general-purpose indexing in English.
- bge-base: Strong semantic performance.
- snowflake-arctic-embed-m-v1.5: High retrieval accuracy with optional dimension reduction for performance tuning.
- multilingual-e5-base: Recommended if your content includes multiple languages.
- Start with a base or small model to establish a performance and relevance baseline.
- Index a representative sample and test retrieval using real queries.
- Use Query Workbench to inspect which chunks are returned for each query.
- Measure indexing time and vector storage to catch early scalability issues.
- Use
dimReductionSizewith supported models to reduce vector size without retraining. - If chunk retrieval is weak, try a different model before adjusting chunking strategy.
2
Vectorize your queries
Just as you needed to vectorize documents, you need to vectorize queries so Lucidworks AI can find the best matching document chunks.Add the LWAI Vectorize Query stage before any stage that alters the query in your query pipeline. Use the same model you selected for the LWAI Chunker Stage in the index pipeline.
LWAI Vectorize Query stage in the correct position in the query pipeline.
3
Configure NHS
Neural Hybrid Search combines and queries to create a flexible balance for any use case.Add the Chunking Neural Hybrid Query stage anywhere between the LWAI Vectorize Query and Solr Query stages. The default configuration is a good starting point, but you must set the Vector Query Field to match the value in the Destination Field Name & Context Output field from the LWAI Chunker Stage.
Chunking Neural Hybrid Query stage in the correct position in the query pipeline.
4
Configure RAG
Add the LWAI Prediction stage anywhere between the LWAI Vectorize Query and Solr Query stages. For this use case, select RAG.
LWAI Prediction stage in the correct position in the query pipeline.
Lucidworks AI supports many use cases. To explore more, see Lucidworks AI use cases.
body and source fields. You can include additional fields if they help improve the responses.If your model requires an API key, add it to the configuration. The key is stored securely.Embedding models for RAG
Embedding models for RAG
Choosing the right embedding model for RAG starts with understanding what you want retrieval to achieve. Embedding models offer different tradeoffs in quality, performance, and language support. The best option depends on your data, your users’ expectations, and your system constraints. Start with a clear goal: Are you trying to surface precise answers from a product catalog? Return relevant support content? Summarize technical documentation?Lucidworks provides a set of pre-trained models that support a wide range of domains. The guidance below focuses on common RAG retrieval needs in B2B commerce, B2C commerce, and knowledge management.Tips for evaluation
RAG in B2B typically supports technical queries against complex product catalogs, configuration rules, and documentation. Accuracy and domain fit matter more than general language quality.
- snowflake-arctic-embed-m-v1.5: High retrieval quality with support for vector size reduction. Recommended starting point for B2B. Optimized for long, structured product content.
- multilingual-e5-base: Strong multilingual support with consistent performance across languages and formats. Good fallback when content spans languages or includes inconsistent structure.
- e5-base-v2: Balanced quality and speed if your data is mostly English and latency is a concern.
- Use real queries from actual users, including edge cases, to measure effectiveness.
- Start with a base or small model and scale up only if needed.
- Inspect retrieved chunks in Query Workbench to verify relevance.
- Reduce vector size using
dimReductionSizein supported models to improve performance. - Switch models easily in the LWAI Prediction stage as your use case evolves.
5
Fine-tune
Everything is set up. Now test your configuration in the Query Workbench. Switch to the JSON view to inspect Neural Hybrid Search results and RAG responses.
RAG responses in the Query Workbench.
- Is the answer backed by retrieved content? The response should only include facts found in the retrieved documents.
- Do citations match the content? References must point to documents that support the answer.
- What happens if nothing useful is retrieved? The system should avoid generating unsupported content.
- Does the answer stay on topic? The response should directly address the query.
- Does the system handle edge cases well? Use ambiguous or off-topic queries to test its behavior.