LangExtract

Set up LangExtract with Ollama and expose it as a local API for PDI.

This setup gives you a local LangExtract API backed by Ollama.

At the end, you will have:

• LangExtract installed in a Python virtual environment
• A FastAPI service running on port 8765
• A test request that returns structured extractions
• A local endpoint ready for use from PDI

These commands assume macOS or Linux.

Use equivalent paths and activation commands on Windows.

Architecture

PDI → HTTP Client → LangExtract API → Ollama model → JSON response

Prerequisites

Component	Version	Notes
Pentaho Data Integration	9.x	Community or EE. Spoon and Pan available.
Python	3.10+	Requires pip and venv.
LangExtract	Current GitHub source	Installed from google/langextract.
FastAPI	Current	REST wrapper for LangExtract.
Uvicorn	Current	ASGI server for the API.
Ollama	0.3+	Local LLM runtime.
Ollama Python package	Current	Required by the sample service code.
Model	llama3.1:8b	Pulled locally with Ollama.
PostgreSQL	Optional	Use only if you plan to persist extracted records.

1

Create project directory

Windows (PowerShell)

mkdir C:\langextract_api
cd C:\langextract_api

Unix

sudo mkdir -p /opt/langextract_api
sudo chown $USER /opt/langextract_api
cd /opt/langextract_api

2

Create and activate virtual environment

Windows

python -m venv langextract_env
langextract_env\Scripts\activate

Unix

python3 -m venv langextract_env
source langextract_env/bin/activate

3

Install dependencies

Identical on both platforms:

pip install fastapi uvicorn langextract ollama

4

Save the script

Save server.py into the project directory:

• Windows: C:\langextract_api\server.py
• Unix: /opt/langextract_api/server.py

Save this file as server.py:

3KB

server.py

Open

5

Pull the Ollama model

Identical on both platforms:

ollama pull llama3.1:8b

6

Start Ollama

Identical on both platforms:

ollama serve

Leave this running in a separate terminal, or configure it as a service (see step 8).

7

Start the LangExtract service

Windows

cd \
cd C:\langextract_api
langextract_env\Scripts\activate
uvicorn server:app --host 0.0.0.0 --port 8765

Unix

cd
cd /opt/langextract_api
source langextract_env/bin/activate
uvicorn server:app --host 0.0.0.0 --port 8765

Run as a background service

Windows - NSSM

# Download nssm from nssm.cc, then:
nssm install LangExtractAPI "C:\langextract_api\langextract_env\Scripts\uvicorn.exe"
nssm set LangExtractAPI AppParameters "server:app --host 0.0.0.0 --port 8765"
nssm set LangExtractAPI AppDirectory "C:\langextract_api"
nssm start LangExtractAPI

Unix - systemd

Create /etc/systemd/system/langextract.service:

[Unit]
Description=LangExtract API Service
After=network.target

[Service]
Type=simple
User=YOUR_USER
WorkingDirectory=/opt/langextract_api
ExecStart=/opt/langextract_api/langextract_env/bin/uvicorn server:app --host 0.0.0.0 --port 8765
Restart=on-failure
RestartSec=5

[Install]
WantedBy=multi-user.target

Then enable and start:

bash

sudo systemctl daemon-reload
sudo systemctl enable langextract
sudo systemctl start langextract
sudo systemctl status langextract

8

Open firewall

Windows

New-NetFirewallRule -DisplayName "LangExtract API" -Direction Inbound -Protocol TCP -LocalPort 8765 -Action Allow

Unix (ufw)

sudo ufw allow 8765/tcp

Unix (firewalld / RHEL)

sudo firewall-cmd --permanent --add-port=8765/tcp
sudo firewall-cmd --reload

9

Verify

Identical on both platforms:

curl http://localhost:8765/health

Response:

StatusCode        : 200
StatusDescription : OK
Content           : {"status":"ok"}
RawContent        : HTTP/1.1 200 OK
                    Content-Length: 15
                    Content-Type: application/json
                    Date: Tue, 07 Apr 2026 09:35:37 GMT
                    Server: uvicorn

                    {"status":"ok"}
Forms             : {}
Headers           : {[Content-Length, 15], [Content-Type, application/json], [Date, Tue, 07 Apr 2026 09:35:37 GMT],
                    [Server, uvicorn]}
Images            : {}
InputFields       : {}
Links             : {}
ParsedHtml        : mshtml.HTMLDocumentClass
RawContentLength  : 15

Swagger UI (browser): http://localhost:8765/docs

---

Verify with a test request

Windows

Run a test extraction:

curl.exe -X POST http://localhost:8765/extract `
  -H "Content-Type: application/json" `
  -d '{\"text\": \"Jane Smith cannot log into the VPN. Error code VPN-403. This is urgent.\", \"prompt\": \"Extract the user, system, issue, urgency, and error code.\", \"examples\": [{\"text\": \"Raj Patel cannot access SAP. Error code ERP-991. Critical issue.\", \"extractions\": [{\"extraction_class\": \"user\", \"extraction_text\": \"Raj Patel\"}, {\"extraction_class\": \"system\", \"extraction_text\": \"SAP\"}, {\"extraction_class\": \"error_code\", \"extraction_text\": \"ERP-991\"}, {\"extraction_class\": \"urgency\", \"extraction_text\": \"Critical\"}]}], \"model_id\": \"llama3.1:8b\"}'

Response:

{"extractions":[{"class":"user","text":"Jane Smith","char_interval":{"start_pos":0,"end_pos":10}},
{"class":"system","text":"VPN","char_interval":{"start_pos":31,"end_pos":34}},
{"class":"issue","text":"cannot log into the VPN","char_interval":{"start_pos":11,"end_pos":34}},
{"class":"urgency","text":"Urgent","char_interval":{"start_pos":64,"end_pos":70}},
{"class":"error_code","text":"VPN-403","char_interval":{"start_pos":47,"end_pos":54}}]}

Unix

Run a test extraction:

curl -X POST http://localhost:8765/extract \
  -H "Content-Type: application/json" \
  -d '{
    "text": "Jane Smith cannot log into the VPN. Error code VPN-403. This is urgent.",
    "prompt": "Extract the user, system, issue, urgency, and error code.",
    "examples": [
      {
        "text": "Raj Patel cannot access SAP. Error code ERP-991. Critical issue.",
        "extractions": [
          {"extraction_class": "user", "extraction_text": "Raj Patel"},
          {"extraction_class": "system", "extraction_text": "SAP"},
          {"extraction_class": "error_code", "extraction_text": "ERP-991"},
          {"extraction_class": "urgency", "extraction_text": "Critical"}
        ]
      }
    ],
    "model_id": "llama3.1:8b"
  }'

Response:

{
  "extractions": [
    {
      "class": "user",
      "text": "Jane Smith",
      "start": 0,
      "end": 10
    }
  ]
}

Success means the API returns an extractions array with extracted values and character offsets.

Call the API from PDI

Use the HTTP Client step:

• URL: http://localhost:8765/extract
• Method: POST
• Content-Type: application/json
• Request body field: your JSON payload
• Response field: response_json

Then parse the response with JSON Input:

• Source is from a field: response_json
• Path: $.extractions[*]
• Fields: class, text, start, end

You can now chain the response into steps such as Select Values, Row Normaliser, or Table Output.

---

PostgreSQL

CREATE SCHEMA IF NOT EXISTS staging;

-- Scenario 1
CREATE TABLE staging.ticket_triage (
  ticket_id     VARCHAR(20) PRIMARY KEY,
  issue_type    VARCHAR(100),
  system        VARCHAR(100),
  urgency       VARCHAR(20),
  reported_by   VARCHAR(100),
  error_code    VARCHAR(100),
  extracted_at  TIMESTAMP DEFAULT NOW()
);

-- Scenario 2
CREATE TABLE staging.patient_extractions (
  id             SERIAL PRIMARY KEY,
  patient_id     VARCHAR(20),
  extract_class  VARCHAR(30),
  extract_text   TEXT,
  char_start     INT,
  char_end       INT,
  extracted_at   TIMESTAMP DEFAULT NOW()
);

-- Scenario 3
CREATE TABLE staging.clause_details (
  id              SERIAL PRIMARY KEY,
  contract_id     VARCHAR(50),
  clause_class    VARCHAR(50),
  clause_text     TEXT,
  char_start      INT,
  char_end        INT,
  extracted_at    TIMESTAMP DEFAULT NOW()
);
CREATE TABLE staging.contract_master (
  contract_id       VARCHAR(50) PRIMARY KEY,
  party_a           TEXT,
  party_b           TEXT,
  effective_date    VARCHAR(100),
  termination       TEXT,
  payment_terms     TEXT,
  liability_cap     TEXT,
  governing_law     TEXT,
  validation_status CHAR(1) DEFAULT 'Y',
  extracted_at      TIMESTAMP DEFAULT NOW()
);

---

Troubleshooting

Common issues:

• ModuleNotFoundError: ollama Install the Python package with pip install ollama.
• Connection refused on port 8765 Confirm that Uvicorn is running.
• Empty or weak extractions Improve the prompt and few-shot examples.
• Model not found Run ollama pull llama3.1:8b again.