Texas tax code assistant with Qdrant, Mistral and OpenAI

Texas tax code assistant with Qdrant, Mistral and OpenAI

An n8n workflow in two parts. The first downloads the Texas Tax Code from the state's public statutes site, cuts it into its
sections and stores each one in a Qdrant vector store, using Mistral embeddings. The second is a chat with an OpenAI agent that
has two tools. One answers a question from the closest sections. The other fetches a section, or a whole chapter, by its number
or name. The agent names the chapter and section it used.

What is inside

• workflow.json: the workflow as n8n exports it, thirty-one nodes and seven sticky notes. It holds no credential.
• README.md: these notes.

What we changed from the original copy

• The state's old download address now returns a web page, not a zip. The copy downloads
  https://tcss.legis.texas.gov/resources/Zips/TX.pdf.zip, the address in the site's own download list, read on 12 November 2026.
  The zip holds 56 chapter PDFs.
• The pattern that cut the text into sections allowed one or two digits on each side of the dot, so it split almost nowhere.
  The other two patterns broke on a section number such as 41A. All three now read \nSec\.\nA[0-9]{1,4}[A-Z]?\.[0-9]{1,5}[A-Z]?\.AA.
  Run on all 56 files with a local copy of n8n's PDF text step, they gave 2,030 sections, with as many labels as texts in every
  file.
• Map To Sections took the chapter from the first PDF for every section, so every section carried the same chapter name. It now
  takes the chapter from the PDF the section came from.
• The chunking node was named for 50,000 characters but cut 30,000, and it rounded its count down. The longest section, 38,277
  characters, lost its end. The node now rounds up, cuts every 30,000 characters in turn and is named Content Chunking @ 30k Chars.
• OpenAI Chat Model set no model, so n8n used its default for that node version, gpt-5-mini, which OpenAI shuts down on 11
  December 2026. It now calls gpt-5.6-terra, the replacement on OpenAI's deprecations page, read on 12 November 2026.
• The example in Search Tool showed a section as Sec 1.01. The stored labels are bare numbers such as 1.001, so the example
  says that.
• A sticky note that invited readers to a chat community and a forum lost that last paragraph.
• Credential links and the webhook ids were removed. Select your own credentials.

Before you start

• An n8n instance, on n8n Cloud or your own server.
• An OpenAI API key and a Mistral API key.
• A Qdrant server that n8n can reach, with its API key if it asks for one. n8n Cloud needs one reachable from the internet.

Set it up

1. In n8n, create a workflow, open its menu and choose Import from File. Pick workflow.json.
2. Add a Mistral Cloud credential and select it in Embeddings Mistral Cloud and Get Mistral Embeddings.
3. Add a Qdrant credential and select it in Qdrant Vector Store, Use Qdrant Search API1 and Use Qdrant Scroll API.
4. In those two HTTP nodes, change http://qdrant:6333 if your Qdrant is somewhere else. The collection name texas_tax_codes
   is in their addresses and in Qdrant Vector Store.
5. Add an OpenAI credential and select it in OpenAI Chat Model.
6. Save the workflow. The two tools call this same workflow by its id, so they work only once it is saved.
7. Click Test workflow to load the code. It runs for a while. Sections go in batches of five with a one second pause, which is
   about 400 batches. Watch the collection fill in Qdrant.
8. Open the chat and ask a question the code answers, such as who must pay the franchise tax. The answer should name a chapter
   and a section.

How the load goes

1. Get Tax Code Zip File downloads the zip, Extract Zip Files unpacks it and Files as Items makes one item per PDF.
   Then Extract PDF Contents reads the text of each one.
2. The text is cut at every section heading in Extract From Chapter, which also builds the list of labels.
   After that, Map To Sections pairs each label with its title and text and adds the chapter name. Next, Sections To List
   makes one item per section and Only Valid Sections drops the empty ones.
3. For Each Section... takes five sections at a time. Long sections are cut into big pieces by Content Chunking @ 30k Chars
   and Split Out Chunks.
4. Qdrant Vector Store embeds and stores them. Recursive Character Text Splitter cuts each piece to 2000 characters, and
   Default Data Loader adds the metadata chapter, section, title and content_order. Then 1sec waits one second and the
   loop goes on.

How the chat goes

When chat message received passes each message to AI Agent, with Window Buffer Memory and OpenAI Chat Model. The agent
has two tools, and each one runs the lower part of this workflow through Execute Workflow Trigger and Switch.

• Ask Tool (query_tax_code_knowledgebase) sends the question to Get Mistral Embeddings, then to Use Qdrant Search API1,
  which asks for the 4 closest points. Get Ask Response returns them as a table of chapter, section, title and content.
• Search Tool (get_tax_code_section) takes a section or a chapter. Use Qdrant Scroll API filters on metadata.section
  when a section is given and on metadata.chapter otherwise, and pages through the matches 100 at a time. Get Search Response
  puts the pieces in content_order and returns them under a short header.

Change the look

• The agent's rules: the system message in AI Agent.
• The source: Get Tax Code Zip File, together with the patterns in Extract From Chapter, which fit this site's layout only.
• How many passages a question reads: the limit in Use Qdrant Search API1.
• The model: OpenAI Chat Model.
• The piece size: the chunk size in Recursive Character Text Splitter.

Good to know

• Running the load again adds every section a second time, since each run stores new points. Delete the collection first.
• Four sections, 151.3185, 155.105, 171.602 and 311.0031, have no separator after their number in the source files. The split
  leaves each inside the section before it.
• content_order is the item's position in the batch. Pieces cut from one 30,000-character chunk share a number, so
  Search Tool can return them in a different order from the one in the law.
• The chapter lookup matches the whole stored name, such as TAX CODE CHAPTER 171. FRANCHISE TAX. The agent can only give it
  after it has seen one.
• The load sends about four million characters to Mistral for embedding, in at least 3,000 pieces. Mistral's rate limits apply,
  and the pause and the batch size are there for them.
• The chat is public once the workflow is active, so anyone with its address can spend your OpenAI and Mistral keys.
• The answers come from a copy of the code loaded on the day you run the workflow. They are not legal advice.
• Questions go to OpenAI and Mistral. The statute text is public.