Stablilizer · API Walkthrough

Interactive API explorer — shared config applies to all tabs.

Developer Docs API Reference Open full endpoint docs, schemas, and example requests.

Base URL

API Key

Health Check

Ready

GET /v1/health

Headers

None required — public endpoint

↓

Waiting

POST /v1/llm-configs

Headers

Content-Type: application/json
Authorization: Bearer <your-api-key>

BYOK (Bring Your Own Key)

Supply your own provider credentials — API key, endpoint, and model

Config name

Extraction model

Used only at extraction time. Compile uses the strength setting below.

Compile base model strength

Set as default LLM config

Request body (edit as needed)

↓

Optimize prompt

Optional

POST /v1/prompt-optimizations

Headers

Content-Type: application/json
Authorization: Bearer <your-api-key>

Uses the prompt, JSON structure, and training data from Step 4.

Primary controls

Number of optimized prompts to produce

How many improved prompt versions to return. Higher values can discover stronger prompts, but increase runtime and cost.

LLM config

Optimization model

Advanced tuning

Training examples to use

How many training examples the optimizer may look at. More examples usually improve coverage, but make the search slower.

Bootstrapped examples limit

How many synthetic example demonstrations the optimizer may create from your training data. Higher values can help when your training set is small, but may overfit or slow the run.

Labeled examples limit

How many original training examples the optimizer may keep as direct reference examples. Higher values preserve more of your dataset, but make the search heavier.

Search rounds

How many search passes the optimizer may run while refining candidate prompts. More rounds can improve results, but each round adds latency.

Candidate prompts to compare

How many competing prompt programs to try before choosing the best one. More candidates widen the search, but also increase cost and runtime.

Retry limit for distinct prompt ideas

How many times to ask for a new variation before giving up on finding another distinct prompt. Higher values help when the model repeats itself, but can spend more tokens.

↓ Request body (live preview — still editable)

↓

Compile a Function

Waiting

POST /v1/functions

Headers

Content-Type: application/json
Authorization: Bearer <your-api-key>

Basic info

Name

Tags (comma-separated)

Description

Scope (who can extract with this function)

Prompt

JSON structure

Training data (JSON array)

[
  {
    "source_text": "The Midnight Market opens at Union Yard on October 3, 2026. Admission is $12. The notice says it is suitable for all ages. No live music is scheduled. The event is supported by Ember Bank.",
    "extracted_json": {
      "event_title": "Midnight Market",
      "start_date": "2026-10-03",
      "ticket_price_usd": 12,
      "audience_rating": "general",
      "has_live_music": false,
      "sponsor_name": "Ember Bank"
    }
  },
  {
    "source_text": "The After Dark Film Expo will be held on November 21, 2026, at the Old Foundry Hall. Tickets are priced at $25. Due to mature screening themes, the listing marks the event for adults only. A live synth set follows the final screening. No sponsor is named in the announcement.",
    "extracted_json": {
      "event_title": "After Dark Film Expo",
      "start_date": "2026-11-21",
      "ticket_price_usd": 25,
      "audience_rating": "adult",
      "has_live_music": true,
      "sponsor_name": null
    }
  },
  {
    "source_text": "Riverfront Makers Day is scheduled for August 9, 2026. Entry costs $7. The flyer says the program is recommended for ages 13 and up. Acoustic performers will play throughout the afternoon. The event is presented by Lantern Mobile.",
    "extracted_json": {
      "event_title": "Riverfront Makers Day",
      "start_date": "2026-08-09",
      "ticket_price_usd": 7,
      "audience_rating": "teen",
      "has_live_music": true,
      "sponsor_name": "Lantern Mobile"
    }
  },
  {
    "source_text": "Skyline Board Game Night starts on December 5, 2026, at Civic Loft. Tickets are $9. The poster says it is open to all ages. A house DJ will play between tournament rounds. The event is backed by Cedar Street Bank.",
    "extracted_json": {
      "event_title": "Skyline Board Game Night",
      "start_date": "2026-12-05",
      "ticket_price_usd": 9,
      "audience_rating": "general",
      "has_live_music": true,
      "sponsor_name": "Cedar Street Bank"
    }
  },
  {
    "source_text": "The Neon Comics Swap lands at Warehouse Annex on July 18, 2026. Admission is $14. The listing recommends the late session for teens. No live music will be part of the program. The event announcement does not mention a sponsor.",
    "extracted_json": {
      "event_title": "Neon Comics Swap",
      "start_date": "2026-07-18",
      "ticket_price_usd": 14,
      "audience_rating": "teen",
      "has_live_music": false,
      "sponsor_name": null
    }
  }
]

Compile mode

This setting decides how the function is compiled, how much work the system does up front, and what kind of extraction strategy is saved for downstream runs.

Prompt variations

Simplest

Compile the prompt directly and ask for a small set of alternative phrasings to work around any LLM model occational weakness.

Best when you want a quick baseline and the task is already clear.

Use optimized prompt(s)

Reuse Step 3

Skip fresh prompt exploration and compile with the optimized prompts generated earlier.

Best when Step 3 already found stronger phrasing and you want to reuse it directly.

Use agents network

Heavy-duty

Build a multi-agent retrieval network instead of one prompt.

Best for reasoning-heavy or dependency-heavy tasks. Also recommended for mission-critical tasks when accuracy is paramount.

Regular compile settings

Number of prompt variations to compile

Optimized prompt settings

Run Step 3 first to generate optimized prompt(s).

Agents network settings

Min % matching fields per training item

0 to 1 (e.g. 0.7 for 70%). A training item counts as passed when at least this fraction of its fields match the expected output.

Min % matching training items overall

0 to 1 (e.g. 0.6 for 60%). The compile is reported as success when at least this fraction of training items pass.

Force agents network to run sequentially (slower but better for complex work)

Run each agents-network step in order instead of batching independent steps in parallel.

Grounding methods

LLM & model

LLM config

Compile base model strength

↓ Request body (live preview — still editable)

↓

Run Extraction

Waiting

POST /v1/extract

Headers

Content-Type: application/json
Authorization: Bearer <your-api-key>

Source text

Minimal run (fastest / cheapest)

Forces the smallest possible run: no grounding methods, a single extraction, and no baseline. Overrides the grounding-method, sample-count, and baseline options below.

LLM override (optional — leave blank to use compiled defaults)

LLM config

Extraction model

Grounding base model strength

Number of extraction samples to produce and process

Grounding methods override (optional)

Override the methods saved at compile time

When off, the run uses the grounding methods chosen during compile. When on, only the methods you check below run (plus the always-on internal guards). Checking none disables every optional method.

Include baseline extraction

Runs the prompt directly once, without Stabilizer scoring or prompt variations, and shows it next to best_result.

↓ Request body (live preview — still editable)

🤖 LLM Configurations GET /v1/llm-configs

Headers

Authorization: Bearer <your-api-key>

Click Refresh to load.

🔍 Inspect / Update / Delete Config

Headers

Content-Type: application/json
Authorization: Bearer <your-api-key>

Config ID

GET /v1/llm-configs/{id}

PATCH body (name / model / provider / is_default)

PATCH /v1/llm-configs/{id} DELETE /v1/llm-configs/{id}

⚙ Functions GET /v1/functions

Headers

Authorization: Bearer <your-api-key>

Click Refresh to load.

🔍 Inspect / Update / Delete Function

Headers

Content-Type: application/json
Authorization: Bearer <your-api-key>

Function ID

GET /v1/functions/{id}

PATCH body (name / description / tags / scope)

PATCH /v1/functions/{id} DELETE /v1/functions/{id}

Quick scope change (sugar for PATCH scope)

Quick grounding strength override (sugar for PATCH grounding_strength — overrides this function's compile strength during extraction; empty = no override)

📄 Recent Extractions GET /v1/extractions

Headers

Authorization: Bearer <your-api-key>

Click Refresh to load.

⏱ Poll Job Status

Headers

Authorization: Bearer <your-api-key>

Job ID

GET /v1/jobs/{id}

📊 Usage GET /v1/usage

Headers

Authorization: Bearer <your-api-key>

From

Click Refresh to load.

🏢 Organization Profile GET /v1/org

Headers

Content-Type: application/json
Authorization: Bearer <your-api-key>

Click Refresh to load.

PATCH org settings

PATCH /v1/org

🔑 API Keys GET /v1/api-keys

Headers

Content-Type: application/json
Authorization: Bearer <your-api-key>

Click Refresh to load.

Revoke key by ID

DELETE /v1/api-keys/{id}

Lifecycle — update budget / suspend / reinstate (marketplace plan changes; requires a full key)

key_id

extract_limit

period

cycle_anchor

PATCH /v1/api-keys/{id}

Create new key

Name

Scope

Budget — optional; leave the limit blank for an unlimited key. 0 = suspended.

extract_limit

period

cycle_anchor (billing_cycle)

Metadata — optional; echoed back in usage.

tenant_id

marketplace_subscription_id

POST /v1/api-keys

📊 Usage & Metering GET /v1/usage

Headers

Authorization: Bearer <your-api-key>

Window — optional half-open [from, to) on created_at. Blank = current period (per-key) / all-time (bulk).

key_id (or "me")

from

GET /v1/api-keys/{id}/usage

GET /v1/usage?group_by=key

✓ Extraction complete — final result