Find beyond the wording.
Use similarity to retrieve passages that relate to the question, even when the phrasing is different.
Search & RAG
Find the context.
Keep the evidence.
Give your application relevant passages, exact matches and the source records behind them. Build answers people can trace back to the original information.
Search & RAG
Find relevant passages. Give your application the sources for its answer.
Illustrative applicationWhat you can build
Make knowledge easier to use.
Keep the exact details.
Use full-text search for names, identifiers and specific terms that a useful answer needs to preserve.
Show where it came from.
Link retrieved IDs to source text, document versions and page references your application can cite.
The application workflow
From a question to grounded context.
Prepare passages once, then choose the retrieval signals that help your application answer its questions.
Suggested record structure
A passage and its source
-
passage_id -
text -
document_id -
page -
version
Your schema, built around your application.
-
Prepare the knowledge
Split documents into passages and store their source information. Configure a table full-text index and an embedding workflow for the fields you need.
-
Retrieve the candidates
Run vector and full-text queries separately. Merge ranked IDs in your application, remove duplicates and load the source rows.
-
Build the answer
Check access, select useful passages and send them to your generation model with their citations. Retain those references in the response.
How it fits together
The data in one place.
The workflow in your hands.
Explore the architecture 
OriginChainDBStore and retrieve.
Stores source records, embeddings and text indexes. Returns relevant matches and the IDs needed to recover their original context.
Coordinate and decide.
Coordinates ingestion, merges retrieval results, checks eligibility and calls the generation model. Owns the prompt, citations and answer evaluation.
Table full-text indexes follow row changes. Dedicated configurations support saved embedders and scheduled embedding runs. Application-supplied vectors and explicit document-index entries need their own update and retry workflow.
Build for real use
Start with a workflow.
Check the whole result.
Can each answer point to a source?
Keep document IDs, locations and versions attached through retrieval and prompt construction. Evaluate using questions with known supporting passages.
What happens when context is missing?
Let the application ask for clarification or decline to answer when the retrieved passages do not support a response.
Does hybrid vector search also run BM25?
The vector hybrid endpoint combines dense and sparse vector rankings. Combining those results with the BM25 text index is a separate application workflow.
Build with OriginChainDB