verbatim-client

Python SDK and CLI for the Verbatim platform — grounded, verbatim answers from curated document collections.

Verbatim is a hosted RAG platform from KR Labs. Every answer is grounded in exact, citable spans from real documents. The platform is organized around collections — curated document sets like the ACL Anthology, or private collections we host for organizations.

This library is the Python client for talking to verbatim.krlabs.eu directly. It wraps the REST API with typed Pydantic responses and ships a verbatim command-line tool.

How it relates to the other Verbatim packages

Install

pip install verbatim-client

Get an API key

1. Sign up at verbatim.krlabs.eu
2. Navigate to API Keys, create a key
3. Export it:

export VERBATIM_API_KEY=vb_your_key_here

SDK usage

Sync

from verbatim_client import VerbatimClient

with VerbatimClient() as client:
    # RAG over the default collection (ACL Anthology)
    result = client.query("What is the attention mechanism in transformers?")
    print(result.answer)
    for cite in result.structured_answer.citations:
        print(f"  [{cite.number}] {cite.text}")

    # Search papers (no LLM, no quota)
    papers = client.search_papers("attention mechanisms", year=2017, limit=5)
    for p in papers:
        print(p.id, p.title)

    # Search papers AND get the chunks that produced the match — useful for
    # feeding directly into a local extractor or your own pipeline. Free.
    papers = client.search_papers(
        "attention mechanisms", limit=5, include_chunks=True,
    )
    for p in papers:
        for chunk in p.matched_chunks or []:
            print(f"  [{chunk.score:.2f}] {chunk.text[:80]}…")

    # Paper detail + BibTeX + full text
    paper = client.get_paper("2017.acl-1.1")
    bibtex = client.get_paper_bibtex("2017.acl-1.1").bibtex
    content = client.get_paper_content("2017.acl-1.1").content

    # Browse facets
    venues = client.facets("venue", q="acl", limit=10)
    for v in venues.items:
        print(v.value, v.count)

    # Cross-collection query (when other collections are available)
    result = client.query(
        "Recent advances in language modeling",
        collection_ids=["anthology", "biorxiv"],
    )

Async

import asyncio
from verbatim_client import AsyncVerbatimClient

async def main():
    async with AsyncVerbatimClient() as client:
        result = await client.query("What is BERT?")
        print(result.answer)

asyncio.run(main())

Streaming

NDJSON streaming exposes documents → highlights → answer stages as they're produced:

with VerbatimClient() as client:
    for event in client.query_stream("What is BERT?"):
        if event["type"] == "documents":
            print(f"Retrieved {len(event['data'])} documents")
        elif event["type"] == "answer" and event.get("done"):
            print("Final answer:", event["data"]["answer"])

Verbatim Transform (bring your own context)

transform is collection-agnostic — pass your own context inline:

result = client.transform(
    "What are the main findings?",
    context=[
        {"content": "Full text of document 1...", "title": "Doc 1"},
        {"content": "Full text of document 2...", "title": "Doc 2"},
    ],
)
print(result.answer)

CLI

The package installs a verbatim command:

# Ask a question
verbatim query "What is BERT?"

# Search papers
verbatim search "attention mechanisms" --year 2017 --limit 5

# Search + include the matched chunks per paper (free, no LLM)
verbatim search "attention mechanisms" --include-chunks --json | jq '.[].matched_chunks'

# List collections
verbatim collections list

# Paper metadata, BibTeX, full text
verbatim paper get 2017.acl-1.1
verbatim paper bibtex 2017.acl-1.1 > vaswani-2017.bib
verbatim paper content 2017.acl-1.1 > vaswani-2017.md

# Facet autocomplete (fuzzy)
verbatim facets author --q vaswani --limit 5

# Verbatim transform over your own context
verbatim transform "What did they find?" --context ./mycontext.json

Every command honors VERBATIM_API_KEY and VERBATIM_API_URL from the environment; override per-invocation with --api-key and --base-url.

Add --json to most commands to get machine-readable output:

verbatim search "attention" --json | jq '.[] | {id, title}'

Configuration

Errors

API errors raise VerbatimError:

from verbatim_client import VerbatimClient, VerbatimError

try:
    client.query("...")
except VerbatimError as e:
    print(e.status_code, e.detail)

Common statuses:

• 401 — authentication required
• 403 — no access to this collection (or non-default collection while the platform's Milvus filter is disabled)
• 404 — collection or paper not found
• 429 — rate limit or quota exceeded
• 451 — legal acceptance required

License

Apache 2.0. See LICENSE.