@jaypie/mcp 0.8.162 → 0.8.164
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/dist/assets.js +3 -3
- package/package.json +1 -1
- package/release-notes/jaypie/1.2.103.md +11 -0
- package/release-notes/llm/1.3.44.md +120 -0
- package/release-notes/llm/1.3.45.md +14 -0
- package/release-notes/mcp/0.8.163.md +17 -0
- package/release-notes/mcp/0.8.164.md +10 -0
- package/release-notes/testkit/1.2.76.md +11 -0
- package/skills/llm.md +70 -1
package/dist/assets.js
CHANGED
|
@@ -55,11 +55,11 @@ function getMcpAssetPaths() {
|
|
|
55
55
|
*/
|
|
56
56
|
function getMcpBuildInfo() {
|
|
57
57
|
return {
|
|
58
|
-
commit: "
|
|
58
|
+
commit: "9c56f9fc"
|
|
59
59
|
,
|
|
60
|
-
version: "0.8.
|
|
60
|
+
version: "0.8.164"
|
|
61
61
|
,
|
|
62
|
-
versionString: "@jaypie/mcp@0.8.
|
|
62
|
+
versionString: "@jaypie/mcp@0.8.164#9c56f9fc"
|
|
63
63
|
,
|
|
64
64
|
};
|
|
65
65
|
}
|
package/package.json
CHANGED
|
@@ -0,0 +1,11 @@
|
|
|
1
|
+
---
|
|
2
|
+
version: 1.2.103
|
|
3
|
+
date: 2026-09-19
|
|
4
|
+
summary: Pulls @jaypie/llm 1.3.44
|
|
5
|
+
---
|
|
6
|
+
|
|
7
|
+
## Changes
|
|
8
|
+
|
|
9
|
+
- **`@jaypie/llm` 1.3.44**: `Llm.ocr` extracts documents to per-page markdown
|
|
10
|
+
over Mistral OCR or the new `llamacloud` provider (LlamaParse), behind one
|
|
11
|
+
request and response shape with fallback across engines.
|
|
@@ -0,0 +1,120 @@
|
|
|
1
|
+
---
|
|
2
|
+
version: 1.3.44
|
|
3
|
+
date: 2026-09-19
|
|
4
|
+
summary: "Llm.ocr extracts documents to per-page markdown over Mistral OCR or LlamaParse behind one shape"
|
|
5
|
+
---
|
|
6
|
+
|
|
7
|
+
## Changes
|
|
8
|
+
|
|
9
|
+
### New Features
|
|
10
|
+
|
|
11
|
+
- `Llm.ocr(document, options?)` is a fifth entry point: a document in,
|
|
12
|
+
per-page markdown out. Static and instance forms mirror `question`,
|
|
13
|
+
including `model` array sugar, `fallback`, `retry`, and `signal`. The
|
|
14
|
+
document (an `https://` URL, a `data:` URI, an S3 key, a local path, or the
|
|
15
|
+
`{ file }` / `{ image }` objects `operate()` accepts) is resolved once before
|
|
16
|
+
the fallback chain runs.
|
|
17
|
+
- New provider `llamacloud` serves LlamaParse over the LlamaCloud Parse API
|
|
18
|
+
v2 with `LLAMA_CLOUD_API_KEY` and match words `llamacloud`, `llamaindex`,
|
|
19
|
+
and `llamaparse`. A parse is a job: submitted by `source_url` or multipart
|
|
20
|
+
upload, polled from one second backing off to five, then fetched. It
|
|
21
|
+
implements `ocr` only; `send` throws `NotImplementedError`. The client is
|
|
22
|
+
`fetch` against four routes; no SDK dependency is taken.
|
|
23
|
+
- The LlamaParse **tier is the model id**: `MODEL.LLAMAPARSE.FAST`,
|
|
24
|
+
`COST_EFFECTIVE`, `AGENTIC` (the provider default), and `AGENTIC_PLUS`.
|
|
25
|
+
A chain such as `[MODEL.MISTRAL.OCR, MODEL.LLAMAPARSE.COST_EFFECTIVE]`
|
|
26
|
+
therefore reads as a plain model list and routes through
|
|
27
|
+
`determineModelProvider`.
|
|
28
|
+
- The API version is pinned per tier in `PROVIDER.LLAMACLOUD.VERSION`; each
|
|
29
|
+
tier accepts its own dated versions and rejects the others (verified live
|
|
30
|
+
2026-09-19). `providerOptions.version` overrides it per call, and the
|
|
31
|
+
response `model` is `<tier id>@<version>`.
|
|
32
|
+
- Common shape: `markdown`, 1-indexed `pages[]` with `header`, `footer`,
|
|
33
|
+
`images`, `success`/`error`, and the vendor page at `raw`; `images[]`;
|
|
34
|
+
`usage` with `pages`, LlamaParse `credits`, and `cost`; `responses[]`; and
|
|
35
|
+
the `provider`/`model`/`fallbackAttempts`/`fallbackUsed` fields `operate`
|
|
36
|
+
carries. Options: `pages` (`[1, 3]` or `"1,3,5-10"`), `tables`
|
|
37
|
+
(`"markdown"` or `"html"`), `images` (download extracted images as `data:`
|
|
38
|
+
URIs), `timeout` (asynchronous job ceiling, default ten minutes), and
|
|
39
|
+
`providerOptions` spread last onto the vendor request.
|
|
40
|
+
- `PAGE_COST` prices document-extraction engines in USD per 1,000 pages,
|
|
41
|
+
keyed by literal id: `mistral-ocr-4-1` at $4 and the four LlamaParse tiers
|
|
42
|
+
at $1.25, $3.75, $12.50, and $56.25 (credits at
|
|
43
|
+
`PROVIDER.LLAMACLOUD.CREDIT_COST`, $1.25 per 1,000). A LlamaParse job's
|
|
44
|
+
recorded credits win over the per-page estimate.
|
|
45
|
+
- `DEFAULT.OCR` names the engine `Llm.ocr` uses when neither a provider nor a
|
|
46
|
+
model is given: Mistral OCR, one synchronous call and the cheapest markdown
|
|
47
|
+
tier on offer.
|
|
48
|
+
- **OCR emulation on chat models**, as `question` emulates: a provider with
|
|
49
|
+
`operate` but no `ocr` transcribes through `src/ocr/OcrEmulator.ts`, so a
|
|
50
|
+
chain such as `[MODEL.MISTRAL.OCR, MODEL.HAIKU]` falls from a native engine
|
|
51
|
+
to any model that accepts files. Each page rides alone in one structured
|
|
52
|
+
`operate()` call (a PDF is trimmed to one page per request, so the model
|
|
53
|
+
never counts or numbers pages), under a fixed `{ confidence, markdown,
|
|
54
|
+
notes }` schema and a verbatim-transcription prompt with `[UNCLEAR: …]` and
|
|
55
|
+
`[ELEMENT: …]` markers. A page scored under 0.4 is transcribed once more
|
|
56
|
+
and the higher score wins. `concurrency` (default 5) bounds pages in
|
|
57
|
+
flight; the first failure aborts the rest and rethrows so the next engine
|
|
58
|
+
starts clean, and unreadable content throws `InternalError` to the same
|
|
59
|
+
end. The response carries `emulated: true`, `pages[].confidence`,
|
|
60
|
+
`usage.tokens`, and `usage.cost` priced from `COST` through the new
|
|
61
|
+
`tokenCost` helper. `user` forwards to the calls. A provider with neither
|
|
62
|
+
(`typesafe`) fails its attempt and the chain moves on.
|
|
63
|
+
- `LlmOcrResponse.emulated` is present on every response (`false` from a
|
|
64
|
+
native engine), matching `LlmQuestionResponse`.
|
|
65
|
+
|
|
66
|
+
### Changed
|
|
67
|
+
|
|
68
|
+
- `MistralProvider.ocr()` takes the common `(document, options)` signature
|
|
69
|
+
and returns the common shape. The provider-specific `{ markdown, model,
|
|
70
|
+
pages, raw }` return shipped in 1.3.43 and is gone: `raw` is
|
|
71
|
+
`responses[0]`, and the 0-indexed `pages[].index` is the 1-indexed
|
|
72
|
+
`pages[].page` with the vendor page at `pages[].raw`. `MistralOcrPage` and
|
|
73
|
+
`MistralOcrResponse` are no longer exported. `OcrRequest` widens to the full
|
|
74
|
+
Mistral `OCRRequest` field set.
|
|
75
|
+
- Client failures on Mistral OCR classify through the chat adapter into the
|
|
76
|
+
typed `LlmError` family, and both engines share one retry loop
|
|
77
|
+
(`src/ocr/runOcrAttempts.ts`), so `retry` and `signal` mean the same thing
|
|
78
|
+
on each. A LlamaParse job that reports `FAILED` throws
|
|
79
|
+
`LlmUnrecoverableError` with the API's message, `CANCELLED` throws
|
|
80
|
+
`LlmAbortError`, and a job that outlives `timeout` throws
|
|
81
|
+
`LlmTransientError` so a chain can try the next engine.
|
|
82
|
+
- `loadFileData` and `getExtensionForMimeType` are exported from the upload
|
|
83
|
+
helpers for reuse; `LlmProvider` gains an optional `ocr`.
|
|
84
|
+
- `extractPdfPages` throws `BadRequestError` on an out-of-range page where it
|
|
85
|
+
threw a vanilla `Error`.
|
|
86
|
+
|
|
87
|
+
### Internal
|
|
88
|
+
|
|
89
|
+
- `MODEL.LLAMAPARSE.*` join `MODEL.MISTRAL.OCR` in `MATRIX_EXCLUDE`: a
|
|
90
|
+
document-extraction engine implements neither `operate()` nor `send()`, so
|
|
91
|
+
every capability cell would fail by construction. They are exercised by
|
|
92
|
+
`tsx test/ocr.ts` (`npm run test:llm:ocr`) and the provider hot specs.
|
|
93
|
+
- Both CI workflows pass `LLAMA_CLOUD_API_KEY` from the
|
|
94
|
+
`CICD_LLAMA_CLOUD_API_KEY` org secret; the LlamaCloud hot spec skips until
|
|
95
|
+
the secret exists.
|
|
96
|
+
|
|
97
|
+
## Testing
|
|
98
|
+
|
|
99
|
+
- Live, 2026-09-19 (`npm run test:llm:ocr`): `page.pdf` and `page.png`
|
|
100
|
+
through Mistral OCR and all four LlamaParse tiers, every run reading the
|
|
101
|
+
fixture text, plus a chain falling from an unknown tier to Mistral. 11 of 11
|
|
102
|
+
passed. Mistral answered in about one second at $0.004 per page; LlamaParse
|
|
103
|
+
completed in 3 to 5 polls (6 to 15 seconds) at 1, 3, 10, and 45 credits per
|
|
104
|
+
page by tier.
|
|
105
|
+
- Live, 2026-09-19, with emulation: the same fixtures through Haiku and
|
|
106
|
+
Luna, plus a chain falling from an unknown tier into Haiku. 15 of 15
|
|
107
|
+
passed. Haiku transcribed a page in 1.5 to 2.5 seconds for $0.0015 to
|
|
108
|
+
$0.0026 (1,200 to 2,200 tokens) at 0.95 confidence; Luna in 1 to 3 seconds
|
|
109
|
+
for $0.0002 to $0.0003 at 0.98 to 0.99. Both marked the fixture's rule as
|
|
110
|
+
`---` or `[ELEMENT: decorative horizontal line]`; Luna read the PDF's
|
|
111
|
+
heading last, where every other engine read it first.
|
|
112
|
+
- New unit coverage for page-selection parsing, document resolution (URL,
|
|
113
|
+
data URI, S3, disk, inline data, PDF trimming), the shared retry loop, the
|
|
114
|
+
LlamaCloud client's routes and status-to-error mapping, the provider's
|
|
115
|
+
submit/poll/fetch cycle (including fast-tier text, failed pages, image
|
|
116
|
+
download, per-tier version, timeout, and abort), the reworked Mistral OCR
|
|
117
|
+
mapping, `Llm.ocr` routing and fallback (native to emulated and back),
|
|
118
|
+
the emulator (page splitting, selection, image and URL input, low-confidence
|
|
119
|
+
re-transcription, concurrency cap, first-failure abort, caller abort, and
|
|
120
|
+
fetch classification), `tokenCost`, and provider routing for every new id.
|
|
@@ -0,0 +1,14 @@
|
|
|
1
|
+
---
|
|
2
|
+
version: 1.3.45
|
|
3
|
+
date: 2026-09-19
|
|
4
|
+
summary: Settle a response the provider cut short as incomplete with the provider's reason
|
|
5
|
+
---
|
|
6
|
+
|
|
7
|
+
## Changes
|
|
8
|
+
|
|
9
|
+
- `operate()` settles `status: "incomplete"` with `error.reason: "incomplete"` (502) when the provider stops the model before it finishes; the partial text stays in `content` and `error.detail` names the provider's reason
|
|
10
|
+
- Every adapter reports it: OpenAI, xAI, and Meta from `status: "incomplete"` plus `incomplete_details.reason` (`max_output_tokens`, `content_filter`); Anthropic from `stop_reason` `max_tokens` or `refusal`; Google from `finishReason` `MAX_TOKENS` or a safety finish; Bedrock from `max_tokens` or a guardrail stop; Fireworks, Mistral, and OpenRouter from `finish_reason` `length` or `content_filter`
|
|
11
|
+
- Previously the loop completed with the truncated text, and a structured request surfaced as prose that failed the format contract ("OCR emulation did not return JSON")
|
|
12
|
+
- `stream()` emits an error chunk from the same stop event on every provider and settles the exchange incomplete; usage from that event still tallies
|
|
13
|
+
- `Llm.ocr` emulation wraps a settled error body in `LlmUnrecoverableError` so a fallback chain logs the provider's reason and moves to the next engine
|
|
14
|
+
- New `LlmResponseErrorReason.Incomplete` and `ParsedResponse.incompleteReason` for adapters
|
|
@@ -0,0 +1,17 @@
|
|
|
1
|
+
---
|
|
2
|
+
version: 0.8.163
|
|
3
|
+
date: 2026-09-19
|
|
4
|
+
summary: Document Llm.ocr and the LlamaCloud provider in the llm skill
|
|
5
|
+
---
|
|
6
|
+
|
|
7
|
+
## Changes
|
|
8
|
+
|
|
9
|
+
### Documentation
|
|
10
|
+
|
|
11
|
+
- `skills/llm.md` gains an "OCR" section: the document forms `Llm.ocr`
|
|
12
|
+
accepts, the page and usage shape, `pages`/`tables`/`images`/`timeout`/
|
|
13
|
+
`providerOptions`, per-tier LlamaParse versions, and fallback across
|
|
14
|
+
engines.
|
|
15
|
+
- The providers table carries a `llamacloud` row, the model constants name
|
|
16
|
+
`MODEL.LLAMAPARSE.*` and `PAGE_COST`, and the environment block carries
|
|
17
|
+
`LLAMA_CLOUD_API_KEY`.
|
|
@@ -0,0 +1,10 @@
|
|
|
1
|
+
---
|
|
2
|
+
version: 0.8.164
|
|
3
|
+
date: 2026-09-19
|
|
4
|
+
summary: Document the incomplete loop stop in the llm skill
|
|
5
|
+
---
|
|
6
|
+
|
|
7
|
+
## Changes
|
|
8
|
+
|
|
9
|
+
- The llm skill's loop stops table gains `incomplete` (502): the provider cut the model off before it finished; `content` holds the partial text and `error.detail` names the provider's reason
|
|
10
|
+
- Add release notes for `@jaypie/llm` 1.3.45
|
|
@@ -0,0 +1,11 @@
|
|
|
1
|
+
---
|
|
2
|
+
version: 1.2.76
|
|
3
|
+
date: 2026-09-19
|
|
4
|
+
summary: Mock Llm.ocr and the LlamaCloud provider
|
|
5
|
+
---
|
|
6
|
+
|
|
7
|
+
## Changes
|
|
8
|
+
|
|
9
|
+
- `Llm.ocr` (static, instance, and on the inner provider) resolves a mock OCR
|
|
10
|
+
response with one page of `_MOCK_OCR_MARKDOWN`.
|
|
11
|
+
- New mocks: `LlamaCloudProvider` and `LlamaCloudClient`.
|
package/skills/llm.md
CHANGED
|
@@ -42,6 +42,7 @@ the alias rather than the id it currently resolves to. Read the id off
|
|
|
42
42
|
| Meta | "muse" (plus the exact name `meta` and a `meta:` prefix) | `MODEL.MUSE_SPARK` |
|
|
43
43
|
| Bedrock | "amazon.nova", "anthropic.claude", "meta.llama", "deepseek.", "google.gemma", "moonshotai.", "openai.gpt-oss", … | `MODEL.NOVA_PRO` |
|
|
44
44
|
| TypeSafe | "jev", "typesafe" | `MODEL.JEV` |
|
|
45
|
+
| LlamaCloud | "llamacloud", "llamaindex", "llamaparse" | `MODEL.LLAMAPARSE.AGENTIC` |
|
|
45
46
|
|
|
46
47
|
The provider name for Gemini models is `"google"` — `"gemini"` is accepted as a deprecated alias.
|
|
47
48
|
|
|
@@ -51,10 +52,12 @@ Mistral's family names mostly do not contain the substring "mistral" (`ministral
|
|
|
51
52
|
|
|
52
53
|
TypeSafe serves System One models (Jev), which answer typed questions and generate no text. The provider implements `question()` only: `operate()`, `send()`, and `stream()` throw. See [Questions](#questions-jev-shape).
|
|
53
54
|
|
|
55
|
+
LlamaCloud serves LlamaParse, which extracts documents and generates no text. The provider implements `ocr()` only. The parse tier is the model id (`MODEL.LLAMAPARSE.FAST`, `COST_EFFECTIVE`, `AGENTIC`, `AGENTIC_PLUS`). See [OCR](#ocr).
|
|
56
|
+
|
|
54
57
|
### Model Constants
|
|
55
58
|
|
|
56
59
|
- **`PROVIDER.<name>.DEFAULT`** — the single default model per provider (above), used when no `model` is given.
|
|
57
|
-
- **`LLM.MODEL.*`** — the named model catalog (e.g. `MODEL.SONNET`, `MODEL.ASTRA`, `MODEL.SOL`, `MODEL.GEMINI_FLASH`, `MODEL.GROK`, `MODEL.MUSE_SPARK`, `MODEL.MUSE_SPARK_CONTRIBUTOR`, `MODEL.NOVA_PRO`, `MODEL.NOVA_LITE`), plus three nested subtrees: `MODEL.FIREWORKS.*` for Fireworks serverless models (`DEEPSEEK`, `GLM`, `GPT_OSS`, `INKLING`, `KIMI`, `MINIMAX`, `NEMOTRON`, `QWEN`), `MODEL.MISTRAL.*` (`LARGE`, `OCR`, `SMALL`), and `MODEL.OPENROUTER.*` for provider-prefixed routes (`GLM`, `LUNA`, `SONNET`). Pick specific models from here. `MODEL.MISTRAL.OCR`
|
|
60
|
+
- **`LLM.MODEL.*`** — the named model catalog (e.g. `MODEL.SONNET`, `MODEL.ASTRA`, `MODEL.SOL`, `MODEL.GEMINI_FLASH`, `MODEL.GROK`, `MODEL.MUSE_SPARK`, `MODEL.MUSE_SPARK_CONTRIBUTOR`, `MODEL.NOVA_PRO`, `MODEL.NOVA_LITE`), plus three nested subtrees: `MODEL.FIREWORKS.*` for Fireworks serverless models (`DEEPSEEK`, `GLM`, `GPT_OSS`, `INKLING`, `KIMI`, `MINIMAX`, `NEMOTRON`, `QWEN`), `MODEL.MISTRAL.*` (`LARGE`, `OCR`, `SMALL`), `MODEL.LLAMAPARSE.*` for LlamaParse tiers (`FAST`, `COST_EFFECTIVE`, `AGENTIC`, `AGENTIC_PLUS`), and `MODEL.OPENROUTER.*` for provider-prefixed routes (`GLM`, `LUNA`, `SONNET`). Pick specific models from here. `MODEL.MISTRAL.OCR` and `MODEL.LLAMAPARSE.*` are document-extraction engines reached through `Llm.ocr`, not chat completions: they are priced per page in `LLM.PAGE_COST` (USD per 1,000 pages, keyed by literal id) and carry no `COST` entry. Amazon's Nova models are first-class ids served over Bedrock; Bedrock's third-party routes are not catalogued — pass the literal id (e.g. `us.anthropic.claude-sonnet-4-6`) and `determineModelProvider` resolves it to `bedrock`.
|
|
58
61
|
- `MODEL.MUSE_SPARK` and `MODEL.MUSE_SPARK_CONTRIBUTOR` are the same Meta model at two tiers. The contributor tier is roughly a tenth of the price because prompts and completions may train Meta models; it is limited to 100 RPM and does not accept `reasoning.effort: "max"`. It is an explicit opt-in: `PROVIDER.META.DEFAULT` is `MODEL.MUSE_SPARK`.
|
|
59
62
|
- The catalog is the **single source of truth for CI coverage**: `packages/llm/test/models.ts` derives the live capability matrix from `MODEL.*` plus each `PROVIDER.*.DEFAULT`, and the workflow shards it by provider. Adding a model to `MODEL.*` puts it under test; no id list exists anywhere else.
|
|
60
63
|
- **Deprecated:** the size-tier map `PROVIDER.<name>.MODEL.{DEFAULT,LARGE,SMALL,TINY}`, the `DEFAULT.MODEL` bundle, and `ALL` are `@deprecated` and retired in 2.0 — use `PROVIDER.*.DEFAULT` for defaults and `MODEL.*` for named models.
|
|
@@ -485,6 +488,70 @@ an attempt with somewhere left to go skips the rate-limit wait. A provider
|
|
|
485
488
|
that can neither `question` nor `operate` counts as a failed attempt and the
|
|
486
489
|
chain moves on.
|
|
487
490
|
|
|
491
|
+
## OCR
|
|
492
|
+
|
|
493
|
+
`Llm.ocr(document, options?)` turns a document into per-page markdown. Two
|
|
494
|
+
engines answer natively behind one request and response shape: Mistral OCR
|
|
495
|
+
(`MODEL.MISTRAL.OCR`, one synchronous call, the default when no provider or
|
|
496
|
+
model is named) and LlamaParse over LlamaCloud (`MODEL.LLAMAPARSE.*`, a job
|
|
497
|
+
that is submitted, polled, and fetched; `LLAMA_CLOUD_API_KEY`). Every other
|
|
498
|
+
provider answers through emulation, as `Llm.question` does: one structured
|
|
499
|
+
`operate()` call per page. A `model` array is therefore a fallback chain
|
|
500
|
+
across native engines and chat models alike.
|
|
501
|
+
|
|
502
|
+
```typescript
|
|
503
|
+
import { Llm, LLM } from "@jaypie/llm";
|
|
504
|
+
|
|
505
|
+
const { emulated, markdown, pages, usage } = await Llm.ocr(
|
|
506
|
+
"./scans/invoice.pdf",
|
|
507
|
+
{
|
|
508
|
+
model: [
|
|
509
|
+
LLM.MODEL.MISTRAL.OCR,
|
|
510
|
+
LLM.MODEL.LLAMAPARSE.COST_EFFECTIVE,
|
|
511
|
+
LLM.MODEL.HAIKU, // emulated: any provider that accepts files
|
|
512
|
+
],
|
|
513
|
+
pages: "1,3-5", // 1-indexed; [1, 3, 4, 5] also works
|
|
514
|
+
tables: "markdown", // or "html"
|
|
515
|
+
},
|
|
516
|
+
);
|
|
517
|
+
|
|
518
|
+
pages[0].page; // 1
|
|
519
|
+
pages[0].markdown;
|
|
520
|
+
pages[0].confidence; // emulated only: the model's self-reported 0..1
|
|
521
|
+
emulated; // true when a chat model transcribed the pages
|
|
522
|
+
usage.pages; // pages billed
|
|
523
|
+
usage.cost; // USD from LLM.PAGE_COST (LlamaParse: from recorded credits; emulated: from COST tokens)
|
|
524
|
+
usage.tokens; // emulated only: every operate() call's usage
|
|
525
|
+
```
|
|
526
|
+
|
|
527
|
+
Emulation sends each page alone (a PDF is trimmed to one page per call, so
|
|
528
|
+
the model never numbers pages), asks for verbatim Markdown with
|
|
529
|
+
`[UNCLEAR: best guess]` and `[ELEMENT: description]` markers, transcribes a
|
|
530
|
+
page scored under 0.4 once more and keeps the higher score, and runs
|
|
531
|
+
`concurrency` pages at a time (default 5); the first failure stops the rest
|
|
532
|
+
so the next engine starts clean. Unreadable content fails the attempt so the
|
|
533
|
+
chain moves on. A provider with native `ocr` (Mistral) always answers
|
|
534
|
+
natively; a provider with neither `ocr` nor `operate` (`typesafe`) fails its
|
|
535
|
+
attempt and the chain moves on.
|
|
536
|
+
|
|
537
|
+
`document` is an `https://` URL (fetched by the vendor), a `data:` URI, an S3
|
|
538
|
+
key (when `CDK_ENV_BUCKET` is set), a local path, or the `{ file, bucket?,
|
|
539
|
+
data?, pages? }` / `{ image, ... }` objects `operate()` accepts. Set
|
|
540
|
+
`images: true` to download extracted images as `data:` URIs onto
|
|
541
|
+
`images[]` and `pages[].images`. `providerOptions` spreads last onto the
|
|
542
|
+
vendor request: Mistral `OCRRequest` fields (`document_annotation_format`,
|
|
543
|
+
`include_blocks`, ...) or LlamaParse `ParseRequestConfiguration` fields
|
|
544
|
+
(`version`, `agentic_options.custom_prompt`, `processing_options`,
|
|
545
|
+
`output_options`, `disable_cache`, ...). `timeout` (default ten minutes)
|
|
546
|
+
bounds a LlamaParse job; expiry throws `LlmTransientError` so a chain can try
|
|
547
|
+
the next engine. `retry` and `signal` behave as on `operate()`.
|
|
548
|
+
|
|
549
|
+
LlamaParse's API version is pinned per tier in `PROVIDER.LLAMACLOUD.VERSION`
|
|
550
|
+
(each tier accepts its own dated versions); the response `model` reads
|
|
551
|
+
`llamaparse-agentic@2026-09-13` so usage records name the configuration. The
|
|
552
|
+
fast tier returns text only, copied into `markdown`. Failed pages arrive with
|
|
553
|
+
`success: false` and `error` rather than failing the call.
|
|
554
|
+
|
|
488
555
|
## Conversation History
|
|
489
556
|
|
|
490
557
|
Continue conversations across calls:
|
|
@@ -734,6 +801,7 @@ OPENROUTER_API_KEY # Required for OpenRouter
|
|
|
734
801
|
XAI_API_KEY # Required for xAI (Grok)
|
|
735
802
|
META_API_KEY # Required for Meta (Muse Spark); MODEL_API_KEY is read as a fallback
|
|
736
803
|
TYPESAFE_API_KEY # Required for TypeSafe (Jev); emulated questions use the answering provider's key
|
|
804
|
+
LLAMA_CLOUD_API_KEY # Required for LlamaParse (Llm.ocr); Mistral OCR uses MISTRAL_API_KEY
|
|
737
805
|
```
|
|
738
806
|
|
|
739
807
|
Keys are resolved via `getEnvSecret()` which supports AWS Secrets Manager.
|
|
@@ -850,6 +918,7 @@ exactly like a provider rate limit.
|
|
|
850
918
|
|----------|--------|---------|
|
|
851
919
|
| `max_turns` | 429 | The model asked for another tool call after `turns` ran out. Nothing failed; the run did not converge. |
|
|
852
920
|
| `tool_errors` | 502 | Tool execution failed six times in a row and the loop stopped. |
|
|
921
|
+
| `incomplete` | 502 | The provider cut the model off before it finished (an output token ceiling, a content filter). `content` holds the partial text; `error.detail` names the provider's reason. |
|
|
853
922
|
|
|
854
923
|
```typescript
|
|
855
924
|
import { LlmResponseErrorReason } from "@jaypie/llm";
|