✨ New
Hybrid Search and Reranker On By Default
New chatbots now use hybrid keyword-plus-vector search for broader recall, and the reranker is enabled by default to order results by relevance before they reach the model. Both settings remain configurable per chatbot.
Customizable Source Chunks (topK)
A new advanced setting lets you control how many source chunks each answer draws from. The default has been increased from 3 to 10 for richer answers, and a slider in the chatbot Advanced section makes the value easy to tune for your content type.
Stable Citation IDs
Citations in answers now carry stable IDs that link back to the exact source chunk. Useful for debugging answer quality and for downstream tools that consume DocuChat output.
Document Title and Author Metadata
Documents now carry title and author metadata that's used by the retriever, so questions like "what did the 2024 annual report say about..." surface the right document more reliably.
🛠️ Improved
- Multi-Group Chatbots: Chatbots with multiple document groups now use every group when answering. Previously only the first group's content was included in the prompt, so questions answerable only by a later group would get "I couldn't find..." responses.
- Reranker Actually Reorders: Fixed a long-standing bug where the reranker was selecting chunks but not actually reordering them for the model. The ranker now controls the final order.
- Refusal Reliability: Chatbots more reliably say they don't know when the answer is not in the uploaded documents, instead of guessing.
- Chunk Boundaries: More natural chunk boundaries for German, Japanese, and Windows-formatted documents.
- Multi-Turn Follow-Ups: Follow-up questions now correctly resolve references to earlier turns in the conversation.
- Re-Upload Cleanup: Re-uploading a document now removes stale chunks from the previous version before indexing the new one.
- Failed Upload Cleanup: Failed uploads no longer leave partial chunks in the search index.
🐛 Fixed
- Embedding Mismatch: Eliminated a systematic mismatch between stored chunk text and its embedding that was quietly degrading retrieval quality.
- DOCX Image Placeholders: DOCX uploads no longer pollute the search index with image placeholder text.
- Image-Only PDFs: Now show a clear error when uploaded instead of importing silently with no searchable content.
- Language Drift: Answers no longer occasionally come back in the wrong language after transient model errors.
- GPT-4 Token Counting: Fixed incorrect token counting for GPT-4 and GPT-4o models.
- Empty-Input Bedrock Calls: Internal Bedrock helpers (language detection, subject generation) no longer error when invoked with an empty message; they short-circuit with a sensible default.