TextLens docs

pip install textlens-ocr
textlens doctor        # check hardware, runtimes and models

from textlens import OCR

ocr = OCR()
result = ocr("invoice.pdf")      # path, URL, bytes, file object, PIL image or numpy array

print(result.text)
print(result.to_markdown())
result.save("output/")             # .json, .md, .txt and tables as .csv

textlens ocr invoice.pdf
textlens ocr invoice.pdf -f markdown -o invoice.md
textlens setup           # optional: detect hardware, choose a profile, install models

result = OCR()("report.pdf")

result.text, result.confidence, result.page_count
for page in result.pages:
    print(page.number, page.provenance.source, page.provenance.model, page.provenance.reasons)
    for block in page.ordered_blocks():
        print(block.type, block.bbox, block.text[:60])

for table in result.tables:
    table.to_csv(), table.to_markdown(), table.to_records()

result.to_markdown(page_markers=True)
result.to_html()
result.to_json()                   # Result.from_json() reads it back
result.save("output/")

import textlens
from textlens import OCR

ocr = OCR()
print(ocr.inspect("report.pdf").summary())    # no rendering, no OCR

ocr("report.pdf", pages="1-3")                  # a page range
OCR(ocr="force")("report.pdf")                # OCR every page
OCR(ocr="off")("report.pdf")                  # native text only

for page in ocr.stream("big.pdf"):             # constant memory
    print(page.number, len(page.text))

doc = textlens.load("big.pdf")                   # lazy: only touched pages run
doc.page(12).text

textlens inspect report.pdf
textlens inspect report.pdf --overlay out/   # draw blocks in reading order
textlens ocr report.pdf --ocr force

OCR(profile="balanced")
OCR(quality="high")                        # shorthand for accurate under auto
OCR(latency_budget_ms=500, memory_budget_gb=4)
OCR(commercial_only=True)                 # skip models with restrictive licenses
OCR(fallback_models=["glm-ocr"], min_confidence=0.85)

for decision in OCR().explain("scan.pdf"):  # dry run: why each model was chosen
    print(decision.explain())

textlens models list                 # catalog + install status
textlens models search table
textlens models info glm-ocr         # model card: hardware, license, limitations
textlens models install glm-ocr
textlens models verify ppocrv6-small # re-check SHA-256
textlens models remove glm-ocr

OCR(model="paddleocr-vl", endpoint="http://gpu-box:8000/v1")

schema = {"vendor": "string", "invoice_number": "string", "date": "date", "total": "float"}
ext = OCR().extract("invoice.pdf", schema=schema)

ext.data                 # {"vendor": ..., "date": ..., "total": ...}
ext.method               # how the values were found
ext.fields["total"]      # found, page, bbox

textlens extract invoice.pdf -s "invoice_number,date,total"

result = OCR()("handbook.pdf")
chunks = result.to_chunks(max_tokens=400, overlap_tokens=40, strategy="semantic")

records = [
    {
        "text": c.text,
        "metadata": {"pages": c.pages, "section": " > ".join(c.section), "boxes": c.bboxes},
    }
    for c in chunks
]

results = OCR().batch("./documents", workers=4)
for r in results:
    if isinstance(r, Exception):
        print("failed:", r)
    else:
        print(r.metadata["source_name"], r.page_count)

textlens batch ./documents -o results/                  # opens a live dashboard
textlens batch ./documents -o results/ --no-dashboard

# pip install "textlens-ocr[anpr]"
from textlens import ANPR

anpr = ANPR(profile="balanced", region="in")
for plate in anpr("car.jpg").plates:
    print(plate.plate, plate.confidence, plate.valid)

pip install "textlens-ocr[server]"
textlens serve                                            # http://127.0.0.1:8000, docs at /docs
TEXTLENS_API_KEYS=k1,k2 textlens serve --host 0.0.0.0 --workers 2 --warmup

curl -H "X-API-Key: k1" -F file=@scan.pdf localhost:8000/ocr
curl -H "X-API-Key: k1" -F file=@scan.pdf -F format=markdown -F pages=1-3 localhost:8000/ocr
curl -H "X-API-Key: k1" -F file=@invoice.pdf -F 'schema={"total":"float","date":"date"}' localhost:8000/extract
curl -H "X-API-Key: k1" -F file=@big.pdf -F async=true localhost:8000/ocr     # → {"job_id": "…"}

pip install "textlens-ocr[mcp]"
textlens mcp --workspace .

{
  "mcpServers": {
    "textlens": {
      "command": "textlens",
      "args": ["mcp", "--workspace", "."]
    }
  }
}

from textlens import OCRBackend, ModelSpec, register_spec

class MyBackend(OCRBackend):
    def _load(self): ...
    def _recognize(self, image, **kwargs): ...   # return a PageOCR

pip install textlens-ocr
textlens setup --profile edge --yes     # download + verify the model, save the profile
textlens ocr sample.jpg

pip install torch --index-url https://download.pytorch.org/whl/cu124
pip install "textlens-ocr[gpu]"
textlens doctor --deep
textlens models install glm-ocr

OCR(profile="balanced")       # PP-OCR for easy pages, GLM-OCR for hard ones
OCR(model="glm-ocr")          # always use GLM-OCR

git clone https://github.com/Srevarshan05/textlens.git && cd textlens
docker build -f deploy/docker/Dockerfile -t textlens .
docker run -p 8000:8000 -e TEXTLENS_API_KEYS=change-me textlens

# before (0.x)
text = OCR(model="glm-ocr").read("scan.png")

# after (2.0)
result = OCR()("scan.png")       # routed; runs without a GPU
text = result.text