<tt>GET /health</tt>

Purpose:

authoritative runtime readiness check
active profile and release identity
safe public telemetry surface for the current room

Important response fields:

status
ready
detail
service_rev
profile_id
router_enabled
max_doc_tokens
default_memory_tokens
default_max_new_tokens
fallback_mode
deterministic_generation
config_class
base_model_class
documents_total
service_uptime_s

Example:

curl -s http://127.0.0.1:8091/health | jq

<tt>GET /documents</tt>

Purpose:

list the documents currently compiled into the active runtime

Response fields per document:

doc_id
name
tokens
chars
memory_tokens
cache_size_mb
compile_time_s
extraction_time_s
total_compile_time_s
source_tokens_total
source_tokens_used
source_truncated
status

Example:

curl -s http://127.0.0.1:8091/documents | jq

<tt>POST /warmup</tt>

Purpose:

explicitly load the runtime before the first compile or query

Example:

curl -s -X POST http://127.0.0.1:8091/warmup | jq '.status, .ready'

<tt>POST /compile_text</tt>

Purpose:

compile already-extracted text directly into LATCH memory

Request fields:

name
text
source_type
memory_tokens optional

Example:

curl -s http://127.0.0.1:8091/compile_text \
  -H 'Content-Type: application/json' \
  -d '{"name":"acme.txt","text":"Acme Corp builds workflow software.","source_type":"text"}'

<tt>POST /compile_file</tt>

Purpose:

upload one file payload and let the runtime extract and compile it

Request fields:

name or filename
content_base64
memory_tokens optional

Supported file types:

.pdf
.txt
.md
.html
.docx
.xlsx
.pptx
.csv
.json
.xml

Example:

curl -s http://127.0.0.1:8091/compile_file \
  -H 'Content-Type: application/json' \
  -d '{"filename":"acme-10k.pdf","content_base64":"<base64>"}'

<tt>POST /query</tt>

Purpose:

run one query against one or more compiled documents

Request fields:

query
doc_ids
memory_tokens optional
max_new_tokens optional

Example:

curl -s http://127.0.0.1:8091/query \
  -H 'Content-Type: application/json' \
  -d '{"query":"Summarize the company in 3 bullets.","doc_ids":["doc_6f3a59b3f8"]}'

<tt>POST /batch</tt>

Purpose:

execute one synchronous batch request against a set of documents and prompts

Use this endpoint when you want one request that evaluates multiple prompts or documents together through the supported batch contract.

This is the customer-facing JSON-in / JSON-out multi-prompt endpoint.

Typical flow:

Compile one or more documents first with POST /compile_text or POST /compile_file.
Save the compiled doc_id values returned by those compile calls.
Put those doc_id values plus a series of prompts into a JSON file.
POST that JSON file to /batch.
Receive one JSON response containing the answers plus optional telemetry and economics blocks.

Key request fields:

session_id
schema_version
documents
queries
options

Per-document fields:

doc_id
filename optional
description optional

Per-query fields:

query_id
prompt
doc_scope
note optional

Execution options:

include_telemetry
include_economics
temperature
max_tokens

Notes:

doc_scope can be a list of doc_id values or the string "all".
temperature must remain 0.0 for the supported deterministic product path.
The response is ordinary JSON, so it is easy to save directly to disk for downstream scripts.

Smallest useful request:

{
  "session_id": "demo-001",
  "schema_version": "1.0",
  "documents": [
    { "doc_id": "doc-10k" }
  ],
  "queries": [
    {
      "query_id": "q001",
      "prompt": "Summarize this company in 2 sentences.",
      "doc_scope": ["doc-10k"]
    }
  ],
  "options": {
    "temperature": 0.0,
    "max_tokens": 256
  }
}

Example batch-request.json:

{
  "session_id": "demo-001",
  "schema_version": "1.0",
  "documents": [
    {
      "doc_id": "doc-10k",
      "filename": "Instructure_10-K.pdf",
      "description": "Primary annual report"
    }
  ],
  "queries": [
    {
      "query_id": "q001",
      "prompt": "What does Instructure do and what markets does it serve?",
      "doc_scope": ["doc-10k"],
      "note": "Opening summary question"
    },
    {
      "query_id": "q002",
      "prompt": "List the main risks in 3 bullets.",
      "doc_scope": "all"
    }
  ],
  "options": {
    "include_telemetry": true,
    "include_economics": true,
    "temperature": 0.0,
    "max_tokens": 512
  }
}

Example:

curl -s http://127.0.0.1:8091/batch \
  -H 'Content-Type: application/json' \
  -d @batch-request.json > batch-response.json

Example response shape:

{
  "session_id": "demo-001",
  "schema_version": "1.0",
  "model": "cdlac_latch_qwen14b_locked_20260317",
  "latch_version": "latch_product_nomount_20260325",
  "generated_at": "2026-03-25T23:45:00Z",
  "session_telemetry": {
    "documents_compiled": 1,
    "total_compile_time_s": 1.24,
    "total_queries": 2,
    "total_wall_clock_s": 2.91,
    "total_tokens_generated": 58
  },
  "economics": {
    "hardware": "NVIDIA A100 80GB PCIe",
    "latch_session_cost_usd": 0.001786,
    "baseline_estimated_session_cost_usd": 0.004911,
    "savings_pct": 63.64,
    "speedup_factor": 2.75
  },
  "queries": [
    {
      "query_id": "q001",
      "prompt": "What does Instructure do and what markets does it serve?",
      "doc_scope": ["doc-10k"],
      "note": "Opening summary question",
      "response": "Instructure provides cloud-based learning software for education and enterprise training.",
      "telemetry": {
        "ttft_s": 0.23,
        "e2e_s": 1.67,
        "tokens_generated": 23,
        "decode_tps": 147.8,
        "decode_mode": "single"
      },
      "baseline_estimate": {
        "estimated_ttft_s": 1.42,
        "estimated_e2e_s": 2.08,
        "ttft_speedup": 6.17,
        "e2e_speedup": 1.25
      }
    }
  ],
  "latch_signature": {
    "license_id": "sha256:1234567890abcdef",
    "model_tuple": "LATCH-Qwen2.5-14B-r1-5k-step3000",
    "image_tag": "latch_product_nomount_20260325"
  }
}

Important response fields:

session_id
schema_version
model
latch_version
generated_at
session_telemetry
economics optional
queries
latch_signature

<tt>POST /documents/clear</tt>

Purpose:

clear the active compiled document set from the runtime

Example:

curl -s -X POST http://127.0.0.1:8091/documents/clear | jq