详细内容或原文请订阅后点击阅览

PDF 的生产 RAG 管道：关系解析、TOC 检索、键入答案

2026年7月7日 16:30 33 Comments
X Twitter Instagram Mail

企业文档智能 [Vol.1 #9A] - 同一篇论文，与第 1 条相同的问题。每块砖一个升级的合同：文档解析、问题解析、检索、生成PDF 的生产 RAG 管道：关系解析、TOC 检索、类型化答案一文首先出现在 Towards Data Science 上。

来源:走向数据科学

III of Enterprise Document Intelligence, a series that builds an enterprise RAG system from four bricks: document parsing, question parsing, retrieval, and generation.

It is the first of two parts on the upgraded pipeline: this part upgrades each brick, one contract at a time, on the same paper and the same question as Article 1 (minimal RAG). The second part,Composing the four RAG bricks into one pipeline, tested on real documents (link to come), wires them into one call and runs it on several real documents.

📓Runnable companion notebooks are on GitHub:doc-intel/notebooks-vol1.

A hundred lines of Python wire four functions together: parse the PDF, parse the question, retrieve a few pages, ask a model.

That pipeline returns the right answer on a clean question against a paper with a built-in table of contents. On a real corpus it breaks the first time the user types“positonal encodig”with two typos, the first time the document is a 200-page contract with no PDF outline, the first time the question asks for every exclusion instead of one, the first time downstream code wants a typed object instead of a string. Four bricks need an upgrade each before the pipeline ships.

  • Document parsing now returns more than a flat line_df: a relational set including a TOC, page-level metadata, and a typed parsing_summary carrying the document’s type, language, and a one-paragraph summary of what it is about.
  • Question parsing turns the noisy user input into a structured brief, with keywords corrected against the corpus’s own vocabulary and an inferred answer shape (single value, listing, table).
  • Retrieval reads the TOC the way an expert would: hand the whole TOC to a small LLM that picks sections by semantic relevance, then merge with the keyword pages.
  • Generation returns a typed answer with one citable span per item, plus four context-quality indicators the pipeline reads to decide whether to ship the answer or run another pass.

    • 1. Where the baseline RAG breaks

    2. Four bricks, upgraded

    four time bricks parsing document first pipeline RAG answer question