Large Engineering Model API

The Large Engineering Model API is a hosted knowledge layer for engineering agents. It returns cited context, source pointers, and workflow hints through an OpenAI-style REST API. It is separate from sim: sim remains the open-source local runtime for operating solver software, while LEM is a network-accessible model/API surface.

The production base URL is:

https://api.svdailab.com/v1

API keys are managed from platform.svdailab.com. Keys use OpenAI-style prefixes such as sk_live_...; product access is controlled by backend scopes.

Model

The first public model is lem-1.

lem-1 is retrieval-first and deterministic. Runtime responses do not call an LLM; they return context blocks, citations, recommended records, and agent hints assembled for the request. Future synthesis models may use the same API family or a successor model id.

Current Beta Scope

lem-1 is currently strongest for COMSOL-style multiphysics modeling questions, especially example discovery, setup patterns, solver troubleshooting, boundary conditions, and cited context packs. It is a retrieval-first context API: it does not run simulations, inspect private project files, or replace solver verification. Broader engineering corpora and client/private corpora may be added under the same API family over time.

Authentication

Send the key in the Authorization header:

curl https://api.svdailab.com/v1/models \
  -H "Authorization: Bearer $SVD_API_KEY"

Keys for lem-1 must have the model:lem-1 scope. Responses include request and rate-limit headers so production agents can log failures and back off when needed.

List Models

GET /v1/models

Example response:

{
  "object": "list",
  "data": [
    {
      "id": "lem-1",
      "object": "model",
      "owned_by": "svd-ai-lab",
      "status": "available",
      "required_scope": "model:lem-1",
      "description": "Retrieval-first Large Engineering Model for engineering agents."
    }
  ]
}

Search Records

POST /v1/search

Use /v1/search when an agent or tool wants raw ranked records, snippets, facets, and citations.

Request fields:

Field	Type	Notes
model	string	Required. Use lem-1.
query	string or content array	Required. Query or task text, limited to 4,000 characters.
max_results	integer	Optional. 1 to 20, default 10.
max_tokens	integer	Optional total snippet budget.
max_tokens_per_record	integer	Optional snippet cap per result.
filters	object	Optional deterministic filters.
include	string array	Optional. snippets, facets, sources, scores.

Supported filters include application_area, source_kind, corpus, module, file_type, and has_local_path.

Example:

curl https://api.svdailab.com/v1/search \
  -H "Authorization: Bearer $SVD_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "lem-1",
    "query": "nonlinear solver convergence for coupled heating",
    "max_results": 5,
    "include": ["snippets", "sources", "facets"]
  }'

Response shape:

{
  "id": "search_...",
  "object": "search",
  "model": "lem-1",
  "results": [
    {
      "id": "record_...",
      "object": "engineering_record",
      "title": "Example record",
      "snippet": "Compact retrieval snippet...",
      "source": {
        "record_id": "record_...",
        "title": "Example record",
        "type": "engineering_reference"
      },
      "score": 8.231
    }
  ],
  "usage": {
    "input_tokens": 12,
    "output_tokens": 160,
    "total_tokens": 172
  },
  "lem": {
    "llm_used": false,
    "mode": "raw_search",
    "result_count": 1
  },
  "request_id": "req_..."
}

Create Response

POST /v1/responses

Use /v1/responses when an agent wants a compact context pack rather than raw search results.

Request fields:

Field	Type	Notes
model	string	Required. Use lem-1.
input	string or content array	Required. Query or task text, limited to 4,000 characters.
max_output_tokens	integer	Optional. 250 to 12,000, default 4,000.
metadata.agent	string	Optional caller name such as codex, claude-code, or chatgpt.
metadata.solver	string	Optional solver hint.
metadata.solver_version	string	Optional solver version hint.
metadata.source_preference	string	Optional. any, local, or online.

Example:

curl https://api.svdailab.com/v1/responses \
  -H "Authorization: Bearer $SVD_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "lem-1",
    "input": "solver convergence issue in a coupled heat-transfer model",
    "max_output_tokens": 900,
    "metadata": {
      "agent": "codex",
      "solver": "comsol",
      "source_preference": "any"
    }
  }'

Response shape:

{
  "id": "resp_...",
  "object": "response",
  "model": "lem-1",
  "output": [
    {
      "type": "message",
      "role": "assistant",
      "content": [
        {
          "type": "output_text",
          "text": "Deterministic context summary..."
        }
      ]
    }
  ],
  "usage": {
    "input_tokens": 14,
    "output_tokens": 426,
    "total_tokens": 440
  },
  "sources": [
    {
      "id": "src_0",
      "record_id": "record_...",
      "title": "Example record",
      "type": "engineering_reference"
    }
  ],
  "lem": {
    "llm_used": false,
    "context_blocks": [],
    "recommended_records": [],
    "agent_hints": []
  }
}

OpenAPI

The versioned OpenAPI document is available at:

GET /v1/openapi.json

Use it for generated clients, agent tool schemas, and integration tests. Access still requires a scoped API key for protected endpoints.

Agent Guidance

Agents should treat sources as citations and lem.context_blocks as retrieval context, not as final engineering truth. For production agent loops, cite returned sources, keep solver actions local and verifiable, and ask the user before loading proprietary project files.

When using LEM with sim, keep the separation clear: LEM provides context through api.svdailab.com; sim operates local solver software on the user’s machine.