After initial filtering, DAKSH applies a context reranker — a shallow attention-based model trained on retrieval-augmented QA pairs — to prioritize the most contextually compatible passages.
Then, the final set of chunks is assembled into a context window, structured as:
cssCopyEdit<Context Start>
[Section 1: Header] …content…
[Section 2: Table: Fee Chart] …content…
[Section 3: Footer Note] …content…
<Context End>
This context is concatenated with the user query and passed into the proprietary LLM for final generation.