Chunking methods and processing settings

Processing settings determine how a document is turned into chunks and what else is written to the index. They are set on the "Document processing" tab in the knowledge base configuration and apply to new documents; files that are already processed have to be processed again.

Where the settings are

The tab shows the embedding model, the "PDF parser" and a note that documents are cut automatically by the recognized headings, paragraphs, lists and tables. Everything else is hidden in the "Advanced settings" block: "Chunking method", the Docling parameters, "Recommended chunk size", "Delimiter for text", "Page rank", "Auto-keyword", "Auto-question", "Excel to HTML", RAPTOR and, for some methods, tag sets and the knowledge graph.

The method of the base is the default. For a single document you can override it in the document list with the "Chunking method" action.

Chunking methods

The list also has methods for emails, audio and resumes (Email, Audio and Resume). If in doubt, start with the General method and check the result in retrieval testing.

Chunk size and delimiter

"Recommended chunk size" sets a threshold in tokens: segments smaller than the threshold are glued together until their sum exceeds it. Small chunks are found more precisely but lose context; large ones blur the similarity to the question and fill the model window faster. For the General method the usual starting point is 512 tokens; for tables and FAQ a row or a pair is already a chunk.

"Delimiter for text" lists the characters on which the text is split into segments before gluing: line breaks, sentence-ending marks. Several characters in one delimiter are given in backticks.

PDF parser

The parameter applies only to PDF; Word is processed separately. The options are:

• Docling (the default): recognizes the page layout, tables and lists. Fields: "Table quality" (fast or accurate mode), "Enable OCR" for scans, "Pages per batch" to manage memory, and "Structure-aware chunking", which keeps lists and headings in one chunk.
• DeepDOC: a built-in visual model for recognizing text, tables and layout. It takes time and resources.
• "Naive": only the text layer, no OCR and no layout analysis. Suitable if all PDFs in the base were created from text; it speeds up processing considerably.
• A provider vision model: a third-party image recognition model, if one is configured in the organization; it is marked as experimental.

Scans without OCR produce empty or garbled chunks; the "Index quality" estimate in the processing progress will show this.

Excel to HTML

With the General method, Excel tables are turned into "header: value" pairs by default. For multi-column tables, merged cells and several tables on one sheet, turn on "Excel to HTML": the sheet is saved as an HTML table, and tables longer than 12 rows are cut into parts of 12 rows. The parameter does not affect PDF.

Auto-keywords and auto-questions

Each chunk is sent to a chat model, which returns N keywords (from 0 to 30) or N questions (from 0 to 10); they are appended to the chunk and improve the hit rate for queries worded differently. Zero turns the feature off. For chunks of about a thousand characters we recommend 3 to 5 keywords and 1 to 2 questions; as the value grows, the benefit falls. Both features increase processing time and token consumption.

RAPTOR

RAPTOR groups chunks by semantic closeness, asks a chat model to briefly retell each group and repeats this level by level until a tree is built, from the original chunks up to an overall summary. The summaries go into the index as additional chunks, so questions that require bringing together several places in a document find the material.

It is turned on with the "Use RAPTOR to enhance retrieval" switch. Parameters: "Prompt" (we recommend leaving it as is), "Max token" per summary, "Threshold" of similarity for joining a cluster (the higher the threshold, the fewer chunks in a cluster), "Max cluster" and "Random seed". The cost is noticeable: memory, time and tokens at every level. Turn it on for bases with multi-step questions and do not turn it on for reference lists and FAQ.

Knowledge graph

The "Extract knowledge graph" switch adds a step in which a model extracts entities and relationships from the chunks and builds a single graph across the whole base. Fields: "Entity types" (by default organization, person, event, category), "Method" ("Light" by default, more economical; "General" on GraphRAG prompts), "Entity resolution" (merging similar entities, costlier) and "Community reports generation" (summaries of entity clusters, costlier).

The graph is updated only when a new document is processed; deleting a file does not rebuild the graph. For an assistant to use the graph, turn on "Use knowledge graph" in its settings. The graph is useful for questions about the links between people, organizations and events, and it noticeably lengthens both processing and the answer.

PageRank

"Page rank" is set for the whole base as an integer from 0 to 100 and adds a constant value to the score of its chunks, so when one assistant searches several bases, the chunks of the base with the higher value come out earlier. Zero means "off". The parameter is supported only with the Elasticsearch search engine; in installations on Infinity it is not applied.