Summary: DocuVision is the analyst tool: a 50-page prospectus becomes searchable passages with page position, then an LLM answers with those passages as context. You do not stitch LangChain, PyPDF, Unstructured, and Chroma. You declare Docs and a Chunks view. The index follows the insert. pip install pixeltable.
The Product#
- Ingest PDFs
- Chunk with position attached
- Ask “What was the Q3 revenue breakdown across regions?” and retrieve the passages
- Synthesize a cited answer from those rows
The longer procedural tutorial is production document RAG. This post is the class-based receipt. Text chunks are the searchable surface. Page images — charts as pixels — are a later column, not a second product we pretend to OCR here.
The Stitch You Delete#
A loader, a splitter, an embedding job, a vector collection, a retriever class, a chain. Delete a PDF and the chunks in Chroma are your problem. In a catalog, delete the document; the view and the index follow. That is the same lineage claim as zombie vectors.
The Receipt#
text and pos come from the splitter. Retrieval is the index. Generation is one Gemini call over those strings — not a second vector database, and not chart OCR. Page images are a later column if you need vision on figures. Same app.py on Cloud: pxt db update → pxt schema update → pxt service update.
What This Is Not#
Not “we read the bar chart as numbers.” That needs page images and a vision column you add later. Not a warehouse. Not LangChain as a religion — you can keep an agent; the document graph does not live in a chain.
People Also Ask#
Why token_limit 300? It is the docs default for passage RAG. Change the iterator kwargs; do not change the table.
Do I need Chroma? No.
How do I go to Cloud? Same file. PIXELTABLE_API_KEY, pxt://org:db, then pxt db update → pxt schema update → pxt service update.
Chunk the File, Not the Stack#
Declare the document. Split it. Index the text. Ask the question. Keep the analyst loop. Delete the loader zoo.
- Pixeltable on GitHub:
pip install pixeltable - ClipFinder — NL video moments
- CallSense — sales-call auditor
- SnapCatalog — shop-the-look
- SafeStream — UGC gate
- ClaimBot — claims triage
- Pixeltable documentation
