Platform settings that decide retrieval quality, cost and data handling in Dify knowledge bases: chunking, index methods, retrieval strategies, metadata, file limits and API access.
Choosing a chunk mode before it is locked: delimiters, maximum length and overlap in general mode, paragraph versus full document parents, the 10,000-token full document limit and which document types suit each.
The cost and accuracy trade-off between embedding-based and keyword-based indexes, why High Quality cannot revert to Economical, and when the keyword index is enough.
Tuning hybrid retrieval in Dify: semantic and keyword weights without a rerank model, when to add a reranker, and why TopK and score threshold only apply in the rerank phase.
Using document metadata fields (string, number, time) to scope retrieval by department, product or date, bulk editing metadata, and the naming and deletion rules that affect every document.
Who can see and use each knowledge base, how creation rights differ from access, and why sensitive document sets need their own knowledge bases and scoped API keys.
Automating document ingestion through the Knowledge API: the 10-key workspace limit, binding keys to specific knowledge bases since 1.17.1, and keeping keys server side.
Indexing images alongside text: automatic extraction from DOCX and XLSX, the attachment limits per chunk, multimodal embedding and rerank models, and extraction plugins for PDFs.
The default 15 MB and 5-file batch limits, how they interact with the Nginx body size, which formats the built-in extractor reads, and settings to change for large document sets.
When question-to-question matching beats paragraph retrieval for FAQ and support content, what Q and A mode generates from each segment, and the extra model cost.
Improving recall with auto-generated chunk summaries and cleaning rules that collapse whitespace or strip URLs and emails, including the cost of a summary model at indexing time.
Crawling public or internal sites into knowledge bases: the crawl services Dify documents, self-hosted crawler options, crawl scope limits and keeping imported pages fresh.
Using retrieval testing to compare chunking, index and rerank settings with real questions before an app goes live, and the signals that point to a chunking problem rather than a model problem.
Tell us what you're working on, whether it's a deployment, an audit, a security test or a cyber range. You'll speak with an engineer who can help you scope it.