model enum Pin the fictional model label and its latency profile. Omit for a per-record mix.
allowed: demo-model-sm | demo-model-md | demo-model-lg
example: model=demo-model-lg
Mock LLM Completion & Streaming Response API
MOCK DATAMock LLM completion fixtures with coherent token usage, tool calls and a replayable delta stream whose chunks concatenate to the exact message.
Generated test data for fixtures and development
- Base URL
- /api/llm-completions
- Category
- Developer Tools
- Capabilities
- 2 routes Seedable
- Last updated
- July 30, 2026
GET /api/llm-completions
Route reference
Code samples Ready-to-copy requests in 4 languages
Parameters Route-specific request inputs 6
| Parameter | Type | Default & allowed | Description |
|---|---|---|---|
model | enum | allowed: demo-model-sm | demo-model-md | demo-model-lg example: demo-model-lg | Pin the fictional model label and its latency profile. Omit for a per-record mix. |
finishReason | enum | allowed: stop | length | tool_calls | content_filter example: length | Pin the terminal reason and therefore the whole coherent shape. Omit for a weighted mix dominated by stop; tool_calls leaves the mix when maxTokens is below the cheapest tool fixture. |
maxTokens | int | default: 64 allowed: 1 – 200 example: 32 | Completion budget in generated word-granular tokens. completionTokens never exceeds it, and finishReason=length always equals it exactly. |
prompt | string | allowed: 1 – 200 example: Summarize this changelog in three bullet points. | Echoed verbatim in every record and counted for promptTokens. Control characters are rejected. It does not influence the completion text or which tool is called. Omit for a per-record fixture prompt. |
cached | boolean | example: true | true forces at least one cached prompt token, false forces exactly zero. Omit for a cache hit on roughly a quarter of records. |
asOf | date | example: 2026-07-30T23:59:59.999Z | Newest possible createdAt. Defaults to the end of the current UTC day and is echoed for replay; createdAt lands within seven days before it. |
finishReason enum Pin the terminal reason and therefore the whole coherent shape. Omit for a weighted mix dominated by stop; tool_calls leaves the mix when maxTokens is below the cheapest tool fixture.
allowed: stop | length | tool_calls | content_filter
example: finishReason=length
maxTokens int Completion budget in generated word-granular tokens. completionTokens never exceeds it, and finishReason=length always equals it exactly.
default: 64
allowed: 1 – 200
example: maxTokens=32
prompt string Echoed verbatim in every record and counted for promptTokens. Control characters are rejected. It does not influence the completion text or which tool is called. Omit for a per-record fixture prompt.
allowed: 1 – 200
example: prompt=Summarize this changelog in three bullet points.
cached boolean true forces at least one cached prompt token, false forces exactly zero. Omit for a cache hit on roughly a quarter of records.
example: cached=true
asOf date Newest possible createdAt. Defaults to the end of the current UTC day and is echoed for replay; createdAt lands within seven days before it.
example: asOf=2026-07-30T23:59:59.999Z
Universal parameters Shared response and formatting options 7
| Parameter | Type | Default & allowed | Description |
|---|---|---|---|
count | int | default: 10 allowed: 1 – 100 example: 3 | How many records to generate (1–100). |
seed | int | example: 42 | Deterministic output: the same seed always returns the same records. Omit for random (the used seed is echoed in meta.seed). |
fields | list | example: id,model | Return only these fields (comma-separated). Mutually exclusive with 'exclude'. |
exclude | list | example: latencyMs | Return all fields except these (comma-separated). |
format | enum | default: json allowed: json | ndjson | csv example: csv | Response format: json envelope, ndjson (one record per line) or csv. |
pretty | boolean | default: false example: true | Pretty-print the JSON response. |
unwrap | boolean | default: false example: true | Drop the envelope: return the raw array/object without data/meta wrapper. |
count int How many records to generate (1–100).
default: 10
allowed: 1 – 100
example: count=3
seed int Deterministic output: the same seed always returns the same records. Omit for random (the used seed is echoed in meta.seed).
example: seed=42
fields list Return only these fields (comma-separated). Mutually exclusive with 'exclude'.
example: fields=id,model
exclude list Return all fields except these (comma-separated).
example: exclude=latencyMs
format enum Response format: json envelope, ndjson (one record per line) or csv.
default: json
allowed: json | ndjson | csv
example: format=csv
pretty boolean Pretty-print the JSON response.
default: false
example: pretty=true
unwrap boolean Drop the envelope: return the raw array/object without data/meta wrapper.
default: false
example: unwrap=true
Response schema Fields returned in each record 9
| Field | Type | Description | Example |
|---|---|---|---|
id | string | Replay handle, shaped cmpl_<completionTokens base36>_<finish code><model code>_<8 lowercase hex>. In ids this route generates the last four hex digits are the record index, which is what keeps them unique within a response; /chunks treats any eight hex digits as a handle, so a hand-written id — like the example here — need not follow that convention. Pass it to /chunks to replay this exact stream. | cmpl_18_s2_5f3ac1d2 |
model | string | Fictional model label; not a real or vendor model, and no model ran. | demo-model-md |
createdAt | string (ISO 8601) | Fictional creation instant, at or before the resolved asOf and within seven days of it. | 2026-07-28T11:42:07.318Z |
prompt | string | The supplied prompt, or a generated fixture prompt. It is not answered by message — no model is involved. | Summarize this changelog in three bullet points. |
message | object | { role: "assistant", content }. content is generic English fixture prose, and an empty string for tool_calls and content_filter. | {"role":"assistant","content":"In practice, the copy is stored next to the source, reads go to the nearest replica. Each field has a default value, the record is versioned on update, the report covers the last seven days. Writes are ordered by a single owner, reads go."} |
finishReason | string | stop, length, tool_calls or content_filter, genuinely derived: length only when the budget was hit, tool_calls only when toolCalls is present. | stop |
toolCalls nullable | object[] | One or two { id, name, arguments } entries, where arguments is a JSON string that parses to an object. Non-null exactly when finishReason is tool_calls. | [{"id":"call_f780b593","name":"lookup_user","arguments":"{\"email\":\"dana@example.com\"}"}] |
usage | object | { promptTokens, completionTokens, cachedPromptTokens, totalTokens }. Generated estimates, not tokenizer output: one token is one whitespace-delimited word, and a tool call costs 2 tokens plus one per 4 characters of arguments. cachedPromptTokens is a subset of promptTokens, never an addition, and totalTokens = promptTokens + completionTokens. | {"promptTokens":7,"completionTokens":44,"cachedPromptTokens":0,"totalTokens":51} |
latencyMs | integer | Fictional wall time for the whole completion: time to the first chunk plus every per-token step. Equals /chunks totalElapsedMs for the same id whenever that id is replayable. | 1194 |
id string Replay handle, shaped cmpl_<completionTokens base36>_<finish code><model code>_<8 lowercase hex>. In ids this route generates the last four hex digits are the record index, which is what keeps them unique within a response; /chunks treats any eight hex digits as a handle, so a hand-written id — like the example here — need not follow that convention. Pass it to /chunks to replay this exact stream.
example: cmpl_18_s2_5f3ac1d2
model string Fictional model label; not a real or vendor model, and no model ran.
example: demo-model-md
createdAt string (ISO 8601) Fictional creation instant, at or before the resolved asOf and within seven days of it.
example: 2026-07-28T11:42:07.318Z
prompt string The supplied prompt, or a generated fixture prompt. It is not answered by message — no model is involved.
example: Summarize this changelog in three bullet points.
message object { role: "assistant", content }. content is generic English fixture prose, and an empty string for tool_calls and content_filter.
example: {"role":"assistant","content":"In practice, the copy is stored next to the source, reads go to the nearest replica. Each field has a default value, the record is versioned on update, the report covers the last seven days. Writes are ordered by a single owner, reads go."}
finishReason string stop, length, tool_calls or content_filter, genuinely derived: length only when the budget was hit, tool_calls only when toolCalls is present.
example: stop
toolCalls object[] nullable One or two { id, name, arguments } entries, where arguments is a JSON string that parses to an object. Non-null exactly when finishReason is tool_calls.
example: [{"id":"call_f780b593","name":"lookup_user","arguments":"{\"email\":\"dana@example.com\"}"}]
usage object { promptTokens, completionTokens, cachedPromptTokens, totalTokens }. Generated estimates, not tokenizer output: one token is one whitespace-delimited word, and a tool call costs 2 tokens plus one per 4 characters of arguments. cachedPromptTokens is a subset of promptTokens, never an addition, and totalTokens = promptTokens + completionTokens.
example: {"promptTokens":7,"completionTokens":44,"cachedPromptTokens":0,"totalTokens":51}
latencyMs integer Fictional wall time for the whole completion: time to the first chunk plus every per-token step. Equals /chunks totalElapsedMs for the same id whenever that id is replayable.
example: 1194
Documented examples
Build-generated requests and complete responses 4 Reproducible mixed completions /api/llm-completions?seed=42&count=10
{
"data": [
{
"id": "cmpl_15_s3_51af0000",
"model": "demo-model-lg",
"createdAt": "2026-07-28T03:37:19.968Z",
"prompt": "Write a commit message for this diff.",
"message": {
"role": "assistant",
"content": "In practice, the flag is read once per session, the response is safe to cache, the job resumes from the last offset. The client backs off between retries, the copy is stored next to the source. Each request carries a trace."
},
"finishReason": "stop",
"toolCalls": null,
"usage": {
"promptTokens": 7,
"completionTokens": 41,
"cachedPromptTokens": 1,
"totalTokens": 48
},
"latencyMs": 1770
},
{
"id": "cmpl_x_s1_86f30001",
"model": "demo-model-sm",
"createdAt": "2026-07-27T07:48:11.022Z",
"prompt": "Propose three names for this feature flag.",
"message": {
"role": "assistant",
"content": "In short, the check runs before every deploy, the report covers the last seven days. The summary is short by design, logs are sampled at a low rate, the summary is short by."
},
"finishReason": "stop",
"toolCalls": null,
"usage": {
"promptTokens": 7,
"completionTokens": 33,
"cachedPromptTokens": 3,
"totalTokens": 40
},
"latencyMs": 547
},
…
"generatedAt": "2026-07-30T15:14:08.000Z"
}
} Truncated at the token budget /api/llm-completions?finishReason=length&maxTokens=32&count=5
{
"data": [
{
"id": "cmpl_w_l2_ad600000",
"model": "demo-model-md",
"createdAt": "2026-07-27T07:05:51.804Z",
"prompt": "Describe what this route returns in one sentence.",
"message": {
"role": "assistant",
"content": "Quick recap, the limit applies per client key, the job resumes from the last offset. Logs are sampled at a low rate, the build fails fast on a type error. The index"
},
"finishReason": "length",
"toolCalls": null,
"usage": {
"promptTokens": 8,
"completionTokens": 32,
"cachedPromptTokens": 0,
"totalTokens": 40
},
"latencyMs": 951
},
{
"id": "cmpl_w_l1_017b0001",
"model": "demo-model-sm",
"createdAt": "2026-07-24T01:49:27.328Z",
"prompt": "Suggest a shorter name for this variable.",
"message": {
"role": "assistant",
"content": "The key point, the flag is read once per session, writes are ordered by a single owner, the worker pool scales with demand. Writes are ordered by a single owner, the header"
},
"finishReason": "length",
"toolCalls": null,
"usage": {
"promptTokens": 7,
"completionTokens": 32,
"cachedPromptTokens": 0,
"totalTokens": 39
},
"latencyMs": 622
},
…
"generatedAt": "2026-07-30T15:14:08.000Z"
}
} Tool-call completions for one prompt /api/llm-completions?finishReason=tool_calls&prompt=What+is+the+weather+in+Berlin%3F&count=5
{
"data": [
{
"id": "cmpl_o_t2_a4cb0000",
"model": "demo-model-md",
"createdAt": "2026-07-25T09:28:28.330Z",
"prompt": "What is the weather in Berlin?",
"message": {
"role": "assistant",
"content": ""
},
"finishReason": "tool_calls",
"toolCalls": [
{
"id": "call_09d5128b",
"name": "create_ticket",
"arguments": "{\"queue\":\"billing\",\"priority\":\"p2\"}"
},
{
"id": "call_b39364cc",
"name": "search_orders",
"arguments": "{\"customerId\":\"cus_7be3c9\",\"status\":\"held\"}"
}
],
"usage": {
"promptTokens": 6,
"completionTokens": 24,
"cachedPromptTokens": 0,
"totalTokens": 30
},
"latencyMs": 907
},
{
"id": "cmpl_b_t1_86650001",
"model": "demo-model-sm",
"createdAt": "2026-07-25T20:45:08.783Z",
"prompt": "What is the weather in Berlin?",
"message": {
"role": "assistant",
"content": ""
…
"generatedAt": "2026-07-30T15:14:08.000Z"
}
} Cache-hit usage on the large model /api/llm-completions?model=demo-model-lg&cached=true&count=20
{
"data": [
{
"id": "cmpl_19_s3_991d0000",
"model": "demo-model-lg",
"createdAt": "2026-07-30T17:32:44.773Z",
"prompt": "Rewrite this comment so it explains why.",
"message": {
"role": "assistant",
"content": "To summarize, the response is safe to cache, the copy is stored next to the source, the id is stable across requests. The parser skips unknown keys, the worker pool scales with demand, the list is sorted by created time. The payload is trimmed to."
},
"finishReason": "stop",
"toolCalls": null,
"usage": {
"promptTokens": 7,
"completionTokens": 45,
"cachedPromptTokens": 4,
"totalTokens": 52
},
"latencyMs": 2323
},
{
"id": "cmpl_8_s3_72710001",
"model": "demo-model-lg",
"createdAt": "2026-07-29T18:57:07.867Z",
"prompt": "Write a commit message for this diff.",
"message": {
"role": "assistant",
"content": "Here is the summary, state is kept out."
},
"finishReason": "stop",
"toolCalls": null,
"usage": {
"promptTokens": 7,
"completionTokens": 8,
"cachedPromptTokens": 7,
"totalTokens": 15
},
"latencyMs": 533
},
…
"generatedAt": "2026-07-30T15:14:08.000Z"
}
} GET /api/llm-completions/chunks
Route reference
Code samples Ready-to-copy requests in 4 languages
Parameters Route-specific request inputs 4
| Parameter | Type | Default & allowed | Description |
|---|---|---|---|
completionId | string | allowed: 5 – 32 example: cmpl_18_s2_5f3ac1d2 | An id from /api/llm-completions, used verbatim as the stream handle and decoded for its token count, finish reason and model — which come back as the completionTokens, finishReason and model fields. Mutually exclusive with tokens, finishReason and model. |
tokens | int | allowed: 1 – 200 example: 24 | Number of content tokens, and therefore of chunks. Rejected together with completionId, which already encodes it. Defaults to 24, and the resolved value is echoed in meta.params whenever no id was supplied. |
finishReason | enum | allowed: stop | length example: length | Terminal reason: stop closes the text with a full stop, length stops mid-sentence with no closing punctuation. Rejected together with completionId, which already encodes it. Defaults to stop, and the resolved value is echoed in meta.params whenever no id was supplied. |
model | enum | allowed: demo-model-sm | demo-model-md | demo-model-lg example: demo-model-lg | Latency profile and record label. Rejected together with completionId, which already encodes it. Defaults to demo-model-md, and the resolved value is echoed in meta.params whenever no id was supplied. |
completionId string An id from /api/llm-completions, used verbatim as the stream handle and decoded for its token count, finish reason and model — which come back as the completionTokens, finishReason and model fields. Mutually exclusive with tokens, finishReason and model.
allowed: 5 – 32
example: completionId=cmpl_18_s2_5f3ac1d2
tokens int Number of content tokens, and therefore of chunks. Rejected together with completionId, which already encodes it. Defaults to 24, and the resolved value is echoed in meta.params whenever no id was supplied.
allowed: 1 – 200
example: tokens=24
finishReason enum Terminal reason: stop closes the text with a full stop, length stops mid-sentence with no closing punctuation. Rejected together with completionId, which already encodes it. Defaults to stop, and the resolved value is echoed in meta.params whenever no id was supplied.
allowed: stop | length
example: finishReason=length
model enum Latency profile and record label. Rejected together with completionId, which already encodes it. Defaults to demo-model-md, and the resolved value is echoed in meta.params whenever no id was supplied.
allowed: demo-model-sm | demo-model-md | demo-model-lg
example: model=demo-model-lg
Universal parameters Shared response and formatting options 4
| Parameter | Type | Default & allowed | Description |
|---|---|---|---|
fields | list | example: completionId,model | Return only these fields (comma-separated). Mutually exclusive with 'exclude'. |
exclude | list | example: chunks | Return all fields except these (comma-separated). |
pretty | boolean | default: false example: true | Pretty-print the JSON response. |
unwrap | boolean | default: false example: true | Drop the envelope: return the raw array/object without data/meta wrapper. |
fields list Return only these fields (comma-separated). Mutually exclusive with 'exclude'.
example: fields=completionId,model
exclude list Return all fields except these (comma-separated).
example: exclude=chunks
pretty boolean Pretty-print the JSON response.
default: false
example: pretty=true
unwrap boolean Drop the envelope: return the raw array/object without data/meta wrapper.
default: false
example: unwrap=true
Response schema Fields returned in each record 9
| Field | Type | Description | Example |
|---|---|---|---|
completionId | string | The replay handle used: the supplied id, or the synthesized one. | cmpl_18_s2_5f3ac1d2 |
model | string | Fictional model label decoded from the id, or the resolved one. | demo-model-md |
finishReason | string | stop or length; also carried on the final chunk. | stop |
content | string | The fully assembled message: exactly chunks.map(c => c.delta).join(""). | In practice, the copy is stored next to the source, reads go to the nearest replica. Each field has a default value, the record is versioned on update, the report covers the last seven days. Writes are ordered by a single owner, reads go. |
chunkCount | integer | Number of chunks in this stream; equals completionTokens. | 44 |
completionTokens | integer | Our word-granular token count for this completion — the same value the list route reports in usage.completionTokens for this id. Prompt-side counts are not encoded in an id, so they are deliberately absent here. | 44 |
timeToFirstChunkMs | integer | chunks[0].elapsedMs; fictional latency, nothing was measured. | 366 |
totalElapsedMs | integer | The last chunk's elapsedMs; equals the list route's latencyMs for this id. | 1194 |
chunks | object[] | Ordered delta records: { index, delta, elapsedMs, finishReason }. delta is one word-granular fragment — the first has no leading space and every later one begins with exactly one space, where a real provider's deltas would be sub-word pieces. elapsedMs is cumulative and strictly increasing. finishReason is null on every chunk but the last. | [{"index":0,"delta":"In","elapsedMs":366,"finishReason":null}] |
completionId string The replay handle used: the supplied id, or the synthesized one.
example: cmpl_18_s2_5f3ac1d2
model string Fictional model label decoded from the id, or the resolved one.
example: demo-model-md
finishReason string stop or length; also carried on the final chunk.
example: stop
content string The fully assembled message: exactly chunks.map(c => c.delta).join("").
example: In practice, the copy is stored next to the source, reads go to the nearest replica. Each field has a default value, the record is versioned on update, the report covers the last seven days. Writes are ordered by a single owner, reads go.
chunkCount integer Number of chunks in this stream; equals completionTokens.
example: 44
completionTokens integer Our word-granular token count for this completion — the same value the list route reports in usage.completionTokens for this id. Prompt-side counts are not encoded in an id, so they are deliberately absent here.
example: 44
timeToFirstChunkMs integer chunks[0].elapsedMs; fictional latency, nothing was measured.
example: 366
totalElapsedMs integer The last chunk's elapsedMs; equals the list route's latencyMs for this id.
example: 1194
chunks object[] Ordered delta records: { index, delta, elapsedMs, finishReason }. delta is one word-granular fragment — the first has no leading space and every later one begins with exactly one space, where a real provider's deltas would be sub-word pieces. elapsedMs is cumulative and strictly increasing. finishReason is null on every chunk but the last.
example: [{"index":0,"delta":"In","elapsedMs":366,"finishReason":null}]
Documented examples
Build-generated requests and complete responses 3 Replay a completion's token stream /api/llm-completions/chunks?completionId=cmpl_18_s2_5f3ac1d2
{
"data": {
"completionId": "cmpl_18_s2_5f3ac1d2",
"model": "demo-model-md",
"finishReason": "stop",
"content": "In practice, the copy is stored next to the source, reads go to the nearest replica. Each field has a default value, the record is versioned on update, the report covers the last seven days. Writes are ordered by a single owner, reads go.",
"chunkCount": 44,
"completionTokens": 44,
"timeToFirstChunkMs": 366,
"totalElapsedMs": 1194,
"chunks": [
{
"index": 0,
"delta": "In",
"elapsedMs": 366,
"finishReason": null
},
{
"index": 1,
"delta": " practice,",
"elapsedMs": 390,
"finishReason": null
},
{
"index": 2,
"delta": " the",
"elapsedMs": 417,
"finishReason": null
},
{
"index": 3,
"delta": " copy",
"elapsedMs": 440,
"finishReason": null
},
{
"index": 4,
"delta": " is",
"elapsedMs": 453,
"finishReason": null
…
"generatedAt": "2026-07-30T15:14:08.000Z"
}
} A 24-chunk stream cut off at the token budget /api/llm-completions/chunks?tokens=24&finishReason=length&model=demo-model-lg
{
"data": {
"completionId": "cmpl_o_l3_e10148e0",
"model": "demo-model-lg",
"finishReason": "length",
"content": "At a high level, the path is resolved at build time, the cache holds warm keys for an hour. The index is rebuilt every",
"chunkCount": 24,
"completionTokens": 24,
"timeToFirstChunkMs": 760,
"totalElapsedMs": 2043,
"chunks": [
{
"index": 0,
"delta": "At",
"elapsedMs": 760,
"finishReason": null
},
{
"index": 1,
"delta": " a",
"elapsedMs": 826,
"finishReason": null
},
{
"index": 2,
"delta": " high",
"elapsedMs": 888,
"finishReason": null
},
{
"index": 3,
"delta": " level,",
"elapsedMs": 942,
"finishReason": null
},
{
"index": 4,
"delta": " the",
"elapsedMs": 985,
"finishReason": null
…
"generatedAt": "2026-07-30T15:14:08.000Z"
}
} Default demo stream /api/llm-completions/chunks
{
"data": {
"completionId": "cmpl_o_s2_f048c398",
"model": "demo-model-md",
"finishReason": "stop",
"content": "At a high level, the summary is short by design, the report covers the last seven days, each field has a default value. The.",
"chunkCount": 24,
"completionTokens": 24,
"timeToFirstChunkMs": 224,
"totalElapsedMs": 832,
"chunks": [
{
"index": 0,
"delta": "At",
"elapsedMs": 224,
"finishReason": null
},
{
"index": 1,
"delta": " a",
"elapsedMs": 256,
"finishReason": null
},
{
"index": 2,
"delta": " high",
"elapsedMs": 283,
"finishReason": null
},
{
"index": 3,
"delta": " level,",
"elapsedMs": 317,
"finishReason": null
},
{
"index": 4,
"delta": " the",
"elapsedMs": 343,
"finishReason": null
…
"generatedAt": "2026-07-30T15:14:08.000Z"
}
} About this API
Coverage & behavior
Generates mock LLM completion records plus a replayable token stream, for tests, streaming interfaces and usage panels. There is no model behind this endpoint: prompt is echoed and counted, and message.content is generic English fixture prose assembled from a seeded word bank. It never answers the prompt, and the tool a tool_calls record picks is unrelated to it.
Token counts are generated estimates, not tokenizer output. One token is one whitespace-delimited word; a tool call costs 2 tokens for its name plus one per 4 characters of arguments. A real sub-word tokenizer splits words into pieces and would report more, so never use these numbers for billing, pricing or context-window math. cachedPromptTokens is a subset of promptTokens, never an addition: totalTokens = promptTokens + completionTokens.
Every record is internally coherent. finishReason=length returns exactly maxTokens completion tokens and text that stops mid-sentence with no closing punctuation; stop returns text that closes its sentence with a full stop (the word bank carries no punctuation of its own, so . is the only terminal character this generator emits) and stays under the budget wherever the budget allows it (only maxTokens=1 forces the two to coincide, so length implies the budget was hit but not the reverse); content_filter returns an empty message and zero completion tokens; tool_calls returns an empty message plus one or two calls whose arguments is a JSON string that parses to an object. toolCalls is non-null exactly when finishReason is tool_calls, and a budget too small for the cheapest tool fixture drops tool_calls from the mix — asking for it explicitly is a 400.
/chunks replays one completion as its ordered delta records. chunks.map(c => c.delta).join("") equals content byte for byte, chunkCount equals completionTokens, and totalElapsedMs equals that record's latencyMs. Any completion's id is the replay handle: it encodes the token count, finish reason and model, and the whole stream is derived from the id string alone, so the same id always rebuilds the same message. To vary a stream by hand, change the eight hex digits at the end of an id.
Called without an id, /chunks synthesizes one from tokens, finishReason and model and returns it in completionId, so you can paste that handle into a later replay. Because an id and those three params are mutually exclusive, meta.params echoes only whichever set was in force — the id on its own, or the resolved trio on its own — and is therefore always a query you can re-issue verbatim.
/chunks is a computed route rather than a generated list, deliberately. A list route would advertise a seed it could not honor, because every chunk record would get its own derived seed and the chunks could not then agree on one message. Instead the whole bounded sequence arrives in one JSON document, and the route reads no clock.
Deliberate limits, so nothing here over-promises. There is no server-sent-events or HTTP streaming: no text/event-stream, no data: frames, no end sentinel. There is no POST route, no request body, no system prompt and no multi-turn input — chat threads live at /api/conversations. The field names are ours; this is not a drop-in for any provider's wire format or SDK, and there are no sampling parameters, no log probabilities, no reasoning tokens and no price fields. There is also no 404: an id is a computed replay handle rather than a key in a finite registry, so a bad id is a 400 rather than a miss — one that quotes the value when the shape is wrong, and names the decoded reason (a tool-call, filtered, empty or over-long completion) when the shape is fine but there is no text to stream.
One honest consequence of a tiny budget: at maxTokens=1 with finishReason=stop every message is a single word, so messages repeat across records even though the records themselves stay distinct through their ids.
Use it for
- Drive a streaming chat interface from a replayable token stream with no model call
- Build fixtures for usage panels that show prompt, cached and completion token counts
- Exercise finish-reason handling for truncated, filtered and tool-call responses
- Reproduce one exact completion in CI by pasting its replay id into a test
Frequently asked questions
How do I mock an LLM streaming response in a test?
Call /api/llm-completions/chunks: it returns the ordered delta records as one JSON document. Concatenate every delta in order to get the message, or feed the records to your own emitter to simulate arrival over time.
Do the deltas reconstruct the completion exactly?
Yes, byte for byte. chunks.map(c => c.delta).join("") equals content, and chunkCount equals completionTokens, because both routes build the stream from the same id-seeded function.
Are the token counts real tokenizer counts?
No. One token is one whitespace-delimited word, and a tool call costs 2 tokens plus one per 4 characters of its arguments. A real sub-word tokenizer would report more, so never use these numbers for billing, pricing or context-window estimates.
Can I point an SDK at this endpoint?
No. It is GET-only JSON with our own field names: there is no POST completion route, no text/event-stream, and no wire-format conformance claim for any provider. Use it as fixture data inside your tests.
Does the completion text answer my prompt?
No. There is no model here. prompt is echoed and counted, and the completion text is generic English fixture prose generated from a seeded word bank.
How do I replay one completion's stream?
Pass that record's id to /chunks. The id is a replay handle that encodes the token count, finish reason and model, so the deltas rebuild that exact message.
Related APIs
Fake Conversation & Chat Data API MOCK DATA
Bounded fake direct, group and support conversations with valid participant references, chronological messages and exact unread counts.
Mock Distributed Trace and Span API MOCK DATA
Generate coherent mock distributed traces and individual spans with bounded timing, parentage, service filters and reproducible failures.
Fake Webhook Delivery API MOCK DATA
Generate coherent webhook delivery histories with safe destinations, bounded retries, status-specific attempts and event-shaped payloads.
CloudEvents 1.0 Mock Data API MOCK DATA
Generate coherent mock CloudEvents 1.0 structured JSON envelopes with all required context attributes and action-specific event data.
Fake API Key & Token Generator MOCK DATA
Clearly fake API keys, JWTs and hex/base64 secrets — structurally valid enough to exercise parsers and middleware, cryptographically meaningless by design.