UNPKG

@juspay/neurolink

Version:

The pipe layer of an AI nervous system: one interface connecting provider neurons to your application, across three inference types — generate, stream and decide. `decide` returns typed, calibrated judgements from a non-generative model (~400ms, ~$0.00002

github.com/juspay/neurolink

juspay/neurolink

153 lines (152 loc) • 8.1 kB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
import type { SpanModelRequest } from "../types/index.js"; export declare const ATTR: { readonly GEN_AI_SYSTEM: "gen_ai.system"; readonly GEN_AI_MODEL: "gen_ai.request.model"; readonly GEN_AI_OPERATION: "gen_ai.operation.name"; readonly GEN_AI_OBSERVATION: "gen_ai.observation.type"; readonly GEN_AI_INPUT_TOKENS: "gen_ai.usage.input_tokens"; readonly GEN_AI_OUTPUT_TOKENS: "gen_ai.usage.output_tokens"; readonly GEN_AI_FINISH_REASON: "gen_ai.response.finish_reason"; readonly GEN_AI_COST_USD: "gen_ai.cost_usd"; readonly GEN_AI_TOOL_NAME: "gen_ai.tool.name"; readonly GEN_AI_TEMPERATURE: "gen_ai.request.temperature"; readonly GEN_AI_MAX_TOKENS: "gen_ai.request.max_tokens"; readonly NL_PROVIDER: "sdk.provider"; readonly NL_MODEL: "sdk.model"; readonly NL_STREAM_MODE: "stream_mode"; readonly NL_TOOL_COUNT: "tool_count"; readonly NL_MESSAGE_COUNT: "message_count"; readonly NL_HAS_TOOLS: "has_tools"; readonly NL_INPUT_LENGTH: "input_length"; readonly NL_OUTPUT_LENGTH: "output_length"; readonly NL_REQUEST_ID: "request_id"; readonly NL_PATH: "path"; readonly NL_HAS_MEMORY: "has_conversation_memory"; readonly NL_COST: "cost"; readonly NL_STRUCTURED_OUTPUT: "structured_output"; readonly NL_HAS_FALLBACK: "has_fallback"; readonly MCP_SERVER_ID: "mcp.server_id"; readonly MCP_TOOL_NAME: "mcp.tool_name"; readonly MCP_TIMEOUT_MS: "mcp.timeout_ms"; readonly MCP_TRANSPORT: "mcp.transport"; readonly MCP_CIRCUIT_STATE: "mcp.circuit_state"; readonly SESSION_ID: "session.id"; readonly USER_ID: "user.id"; readonly MEMORY_TYPE: "memory.type"; readonly MESSAGE_COUNT: "message.count"; readonly CONTENT_LENGTH: "content.length"; readonly RAG_FILE_COUNT: "rag.file_count"; readonly RAG_STRATEGY: "rag.strategy"; readonly RAG_CHUNK_SIZE: "rag.chunk_size"; readonly RAG_TOP_K: "rag.top_k"; readonly RAG_RESULT_COUNT: "rag.result_count"; readonly MSG_COUNT: "message.build.count"; readonly MSG_HAS_IMAGES: "message.has_images"; readonly MSG_HAS_FILES: "message.has_files"; readonly MSG_HAS_SYSTEM_PROMPT: "message.has_system_prompt"; readonly MSG_TOTAL_CONTENT_LENGTH: "message.total_content_length"; readonly MSG_IS_MULTIMODAL: "message.is_multimodal"; readonly FILE_NAME: "file.name"; readonly FILE_MIMETYPE: "file.mimetype"; readonly FILE_SIZE_BYTES: "file.size_bytes"; readonly FILE_CATEGORY: "file.category"; readonly FILE_OUTPUT_LENGTH: "file.output_length"; readonly FILE_SUCCESS: "file.success"; readonly FILE_PROCESSOR_USED: "file.processor_used"; readonly FILE_CONFIDENCE: "file.detection_confidence"; readonly FILE_HAS_IMAGES: "file.has_images"; readonly FILE_IMAGE_COUNT: "file.image_count"; readonly FILE_ERROR: "file.error"; readonly FILE_SKIPPED_REASON: "file.skipped_reason"; readonly FILE_TOTAL_COUNT: "file.total_count"; readonly FILE_INCLUDED_COUNT: "file.included_count"; readonly FILE_CONTENT_TYPE: "file.content_type"; readonly VIDEO_DURATION_SEC: "video.duration_sec"; readonly VIDEO_WIDTH: "video.width"; readonly VIDEO_HEIGHT: "video.height"; readonly VIDEO_CODEC: "video.codec"; readonly VIDEO_AUDIO_CODEC: "video.audio_codec"; readonly VIDEO_FORMAT: "video.format"; readonly VIDEO_BITRATE: "video.bitrate"; readonly VIDEO_KEYFRAMES_EXTRACTED: "video.keyframes_extracted"; readonly VIDEO_TARGET_FRAMES: "video.target_frames"; readonly VIDEO_INTERVAL_SEC: "video.interval_sec"; readonly VIDEO_FRAME_QUALITY: "video.frame_quality"; readonly VIDEO_FRAME_RESIZE_PX: "video.frame_resize_px"; readonly VIDEO_FRAMES_RESIZED: "video.frames_resized"; readonly VIDEO_TOTAL_FRAME_BYTES: "video.total_frame_bytes"; readonly VIDEO_HAS_SUBTITLES: "video.has_subtitles"; readonly VIDEO_SUBTITLE_LENGTH: "video.subtitle_length"; readonly VIDEO_TEXT_CONTENT_LENGTH: "video.text_content_length"; readonly VIDEO_HAS_METADATA: "video.has_metadata"; readonly VIDEO_HAS_KEYFRAMES: "video.has_keyframes"; readonly VIDEO_TEMP_PATH: "video.temp_path"; readonly VIDEO_BYTES_WRITTEN: "video.bytes_written"; readonly VIDEO_WRITE_DURATION_MS: "video.write_duration_ms"; readonly VIDEO_STEP: "video.step"; readonly CONTEXT_STAGE: "context.compaction_stage"; readonly CONTEXT_TOKENS_BEFORE: "context.tokens_before"; readonly CONTEXT_TOKENS_AFTER: "context.tokens_after"; readonly MW_COUNT: "middleware.count"; readonly MW_NAMES: "middleware.names"; readonly AR_TAG: "autoresearch.tag"; readonly AR_BRANCH: "autoresearch.branch"; readonly AR_PHASE: "autoresearch.phase"; readonly AR_PHASE_FROM: "autoresearch.phase_from"; readonly AR_PHASE_TO: "autoresearch.phase_to"; readonly AR_RUN_COUNT: "autoresearch.run_count"; readonly AR_KEEP_COUNT: "autoresearch.keep_count"; readonly AR_STATUS: "autoresearch.status"; readonly AR_METRIC: "autoresearch.metric"; readonly AR_BEST_METRIC: "autoresearch.best_metric"; readonly AR_DIRECTION: "autoresearch.metric_direction"; readonly AR_COMMIT: "autoresearch.commit"; readonly AR_DURATION_MS: "autoresearch.duration_ms"; readonly AR_DESCRIPTION: "autoresearch.description"; readonly AR_ERROR_CODE: "autoresearch.error_code"; }; /** Default ceiling for serialized span attribute values. */ export declare const SPAN_ATTRIBUTE_MAX_CHARS = 40000; /** * The ceiling in force: `NEUROLINK_SPAN_ATTRIBUTE_MAX_CHARS` when it is a * positive whole number (raised to the minimum if lower), otherwise the * default. Read on every call, so a change applies to the next span. */ export declare function spanAttributeMaxChars(): number; /** * @description Returns a copy of a value with credentials masked: values under keys such as `access_token`, `password`, `authorization` or `set-cookie`, bearer tokens, and `name=value` credentials inside text. Purely numeric values after a credential-looking name (e.g. `max_tokens: 4096`) are left alone. * @param value {JsonValue | object} - Content to redact; objects and arrays are walked recursively. * @returns {JsonValue | object} A redacted copy, serializing exactly as the input when nothing in it is sensitive. * @throws {Error} When the value cannot be walked; tracing callers use `spanContentAttribute`, which turns that into an omitted value. */ export declare function redactSpanContent(value: unknown): unknown; /** * `spanJsonAttribute` for content a person or model wrote or a tool returned * (arguments, results, answers), with credentials masked first. Use the plain * form for metadata such as usage counts and server ids. * * A structured value over the cap is shortened by cutting every string to a * per-string budget in one walk, tightening the budget only if the result * still does not fit, so the attribute stays valid JSON instead of ending * mid-token. */ export declare function spanContentAttribute(value: unknown, maxChars?: number): string; /** * Serialize an arbitrary value for a span attribute, hard-capped at * `maxChars` so a pathological prompt or tool result can't put megabytes * on a single span. Strings pass through unserialized; everything else is * JSON-stringified with a String() fallback for circular structures. */ export declare function spanJsonAttribute(value: unknown, maxChars?: number): string; /** * Serialize a model request for a span attribute without losing the * conversation to the cap. * * A head cut keeps the system prompt and drops everything after it, so a long * system prompt hides every message, the user's question included. Here the * prompt, the current question and this turn's newest messages are budgeted * first, earlier messages are dropped whole and counted, and the system prompt * gets whatever room is left. A tool result is never kept without the call it * answers. The result is always valid JSON. */ export declare function spanModelRequestAttribute(rawRequest: SpanModelRequest, maxChars?: number): string;