Sarvam Vision

Sarvam Vision is a 3B parameter state-space Vision Language Model (VLM) purpose-built for high-accuracy Document Intelligence. It powers our Document Intelligence pipeline.

At a Glance

Why Sarvam Vision?

One of the most challenging problems in vision AI today is accurate document intelligence for Indian languages. Much of India’s knowledge—historical texts, government records, academic papers, and cultural archives—remains locked in libraries, scanned collections, and legacy documents. Unlocking this vast repository is essential for preserving cultural heritage and making knowledge accessible.

While frontier Vision Language Models have set a high bar for processing modern English documents, a significant gap remains: most global models treat Indian languages as secondary, often resulting in lower accuracy for regional scripts. Sarvam Vision bridges this gap with native support for 22 Indian languages, delivering world-class accuracy where others fall short.

What You Can Do

• Text Extraction: Extract text from PDFs and scanned documents in 23 languages (22 Indian + English)
• Tables: Convert complex tables to HTML or Markdown
• Structure Preservation: Maintain document layout, reading order, and hierarchies

Supported Languages

All 22 official Indian languages plus English:

Capabilities

Quick Start

Get started with Document AI for high-fidelity text extraction across all supported languages. Create the job in one call, poll until it reaches a terminal state, then fetch the output.

1
import os, time
2
from sarvamai import SarvamAI
3
4
client = SarvamAI(api_subscription_key=os.environ["SARVAM_API_KEY"])
5
6
# Create and submit a digitise job in one call
7
with open("document.pdf", "rb") as f:
8
    job = client.doc_ai.digitise(
9
        file=[("document.pdf", f, "application/pdf")],
10
        language="hi-IN",
11
        output_format="md",   # "html" (default) or "md"
12
    )
13
print("job created:", job.job_id, "| status:", job.status)
14
15
# Poll until the job reaches a terminal state
16
TERMINAL = {"completed", "partially_completed", "failed", "rejected"}
17
while True:
18
    st = client.doc_ai.get_status(job_id=job.job_id)
19
    print(f"status={st.status} pages={st.usage.pages_processed}/{st.usage.pages_total}")
20
    if st.status.lower() in TERMINAL:
21
        break
22
    time.sleep(5)
23
24
# Fetch the output (a ZIP holding the .md plus page-level JSON)
25
if st.status.lower() in ("completed", "partially_completed"):
26
    dl = client.doc_ai.get_download_url(job_id=job.job_id)
27
    print("download:", dl.method, dl.url)

See the Document AI overview for schema-based Extract, output formats, and the full job lifecycle.

1
from sarvamai import SarvamAI
2
3
client = SarvamAI(
4
    api_subscription_key="YOUR_SARVAM_API_KEY"
5
)
6
7
# Create a document intelligence job
8
job = client.document_intelligence.create_job(
9
    language="hi-IN",
10
    output_format="md"
11
)
12
print(f"Job created: {job.job_id}")
13
14
# Upload document
15
job.upload_file("document.pdf")
16
print("File uploaded")
17
18
# Start processing
19
job.start()
20
print("Job started")
21
22
# Wait for completion
23
status = job.wait_until_complete()
24
print(f"Job completed with state: {status.job_state}")
25
26
# Get processing metrics
27
metrics = job.get_page_metrics()
28
print(f"Page metrics: {metrics}")
29
30
# Download output (ZIP file containing the processed document)
31
job.download_output("./output.zip")
32
print("Output saved to ./output.zip")

Model Specifications

Limits

Next Steps