Pixeltable for document-heavy insurance

FNOL and claim-file pipeline

1. Insert claim_id, doc_kind (fnol, estimate, medical, photo), and the PDF or image.
2. Split PDFs with document_splitter (sentence,token_limit).
3. Index passage text. Index photos with CLIP when the file is an image.
4. Retrieve passages for a desk question, filtered by claim_id.
5. Generate a summary from those passages only. Keep the prompt so a reviewer can re-run it.

Worked example: retrieve FNOL passages for one claim, then summarize

The prompt tells the model to use only retrieved passages. That is a grounding instruction, not a guarantee against hallucination. Keep the claim file in S3 if that bucket is the archive. Keep pgvector if Postgres is already the claims store and you only need ANN. Lineage of which model produced which summary is the computed column on that row; Pixeltable versions the table when you apply a new expression.

import pixeltable as pxt
from pixeltable.functions.document import document_splitter
from pixeltable.functions.huggingface import sentence_transformer
from pixeltable.functions.openai import chat_completions

TableModel = pxt.model_base()
text_embed = sentence_transformer.using(model_id='all-MiniLM-L6-v2')

class ClaimDocs(TableModel, name='claim_docs'):
    claim_id: pxt.String
    doc_kind: pxt.String
    document: pxt.Document

class Passages(
    TableModel,
    name='passages',
    base=ClaimDocs,
    iterator=document_splitter(
        document=ClaimDocs.document, separators='sentence,token_limit', limit=300
    ),
):
    __indexes__ = [
        pxt.EmbeddingIndex(text, string_embed=text_embed),
    ]

@pxt.query
def retrieve_claim(question: str, claim_id: str, limit: int = 8):
    sim = Passages.text.similarity(string=question)
    return (
        Passages.where(ClaimDocs.claim_id == claim_id)
        .order_by(sim, asc=False)
        .limit(limit)
        .select(
            doc_kind=ClaimDocs.doc_kind,
            text=Passages.text,
            score=sim,
        )
    )

@pxt.udf
def build_claim_prompt(question: str, passages: list[dict]) -> str:
    lines = '\n'.join(f"[{row['doc_kind']}]: {row['text']}" for row in passages)
    return (
        'Answer using only the retrieved passages. If they do not contain the answer, say so.\n\n'
        f'{lines}\n\n'
        f'Question: {question}'
    )

class DeskQuestions(TableModel, name='desk_questions'):
    claim_id: pxt.String
    question: pxt.String
    passages = retrieve_claim(question, claim_id, limit=8)
    prompt = build_claim_prompt(question, passages)
    answer = chat_completions(
        messages=[{'role': 'user', 'content': prompt}],
        model='gpt-4o-mini',
    ).choices[0].message.content

# pxt init
# pxt schema update app.py claims -f

To run a sketch, install pixeltable[serve] and the packages its model functions import (such as sentence-transformers, openai, or openai-whisper), export the provider key a model call reads (such as OPENAI_API_KEY), save the file as app.py, and run the commands in its last lines. The tables start empty: insert a row, or POST to an insert route, then select the computed columns. Get started covers the CLI.

Honest alternatives per slot

Where Pixeltable fits

Pixeltable is the claim-file table: PDFs, a splitter view, an embedding index, retrieval filtered by claim_id, and an LLM column whose prompt is the retrieved passages. Keep the policy admin system. Keep S3 as the archive. Keep Unstructured or LlamaParse in front of insert when layout needs a specialist parser. Keep pgvector or Qdrant when that engine is already the index. This page does not claim a compliance certification. The architecture is /reference-architecture/multimodal-rag. The how-tos are multimodal RAG and chunk PDFs for RAG.

Questions