From 7fb49109398b74a3da2d6eb70fdf4791eb8c165e Mon Sep 17 00:00:00 2001 From: "detail-app[bot]" <180357370+detail-app[bot]@users.noreply.github.com> Date: Wed, 29 Jul 2026 16:12:45 +0000 Subject: [PATCH] docs: correct hybrid prefetch-limit and rag search-top-k defaults --- docs/configuration.md | 2 +- docs/retrieval-pipeline.md | 4 ++-- 2 files changed, 3 insertions(+), 3 deletions(-) diff --git a/docs/configuration.md b/docs/configuration.md index c05e152d..60916c09 100644 --- a/docs/configuration.md +++ b/docs/configuration.md @@ -106,7 +106,7 @@ See [pipeline-commands.md](pipeline-commands.md#hybrid-qdrant-setup) for the col - `QDRANT_COLLECTION_PDFS` (default `java-chat-qwen3-embedding-4b-2560-pdfs`) - `QDRANT_DENSE_VECTOR_NAME` (default `dense`) — named vector for dense embeddings - `QDRANT_SPARSE_VECTOR_NAME` (default `bm25`) — named vector for BM25 sparse tokens -- `HYBRID_PREFETCH_LIMIT` (default `20`) — per-stage prefetch limit for RRF fusion queries +- `HYBRID_PREFETCH_LIMIT` (default `14`) — per-stage prefetch limit for RRF fusion queries - `HYBRID_RRF_K` (default `60`) — reciprocal-rank-fusion k parameter used by Qdrant query fusion - `HYBRID_QUERY_TIMEOUT` (default `10s`) — timeout for hybrid search queries - `APP_QDRANT_ENSURE_PAYLOAD_INDEXES` (default `true`) — create payload indexes on startup diff --git a/docs/retrieval-pipeline.md b/docs/retrieval-pipeline.md index ad748fad..b6885258 100644 --- a/docs/retrieval-pipeline.md +++ b/docs/retrieval-pipeline.md @@ -275,7 +275,7 @@ All properties are bound via `AppProperties` (`@ConfigurationProperties(prefix = |---|---|---| | `app.qdrant.dense-vector-name` | `dense` | Named vector key for dense embeddings | | `app.qdrant.sparse-vector-name` | `bm25` | Named vector key for sparse BM25 tokens | -| `app.qdrant.prefetch-limit` | `20` | Per-stage candidate count for each dense/sparse prefetch before RRF fusion | +| `app.qdrant.prefetch-limit` | `14` | Per-stage candidate count for each dense/sparse prefetch before RRF fusion | | `app.qdrant.rrf-k` | `60` | RRF k parameter: `score = Sum(1 / (k + rank))` | | `app.qdrant.query-timeout` | `10s` | Timeout for hybrid search fan-out across all collections | | `app.qdrant.ensure-payload-indexes` | `true` | Create payload indexes on startup for metadata filtering | @@ -296,7 +296,7 @@ All four must be non-blank and distinct (validated on startup). | Property | Default | Validation | Description | |---|---|---|---| -| `app.rag.search-top-k` | `12` | Must be > 0 | Candidates fetched from hybrid search before reranking | +| `app.rag.search-top-k` | `8` | Must be > 0 | Candidates fetched from hybrid search before reranking | | `app.rag.search-return-k` | `6` | Must be > 0, must be <= `search-top-k` | Results returned to the LLM after reranking | | `app.rag.reranker-timeout` | `8s` | Must be positive and less than `20s` | Timeout for LLM reranking call | | `app.rag.search-citations` | `3` | Must be >= 0 | Citation references included in the response |