@arbidocs/client 0.5.122 → 0.5.123

package/dist/index.d.cts

@@ -9694,6 +9694,7 @@ interface components {
              *     5. Prefer digest=true from the outset. It is the normal retrieval mode: findings with citations, compact context. Decide against it only when the task needs the exact words — translation, quotation, close reading of style or drafting. Choosing digest=false IS that decision, and it is final for that document: digest is a way of retrieving, not a way of processing text you already hold, so once you have loaded a document in full a digest of it is refused. Reason over the text you have.
              *     6. Open-ended investigative queries ("find interesting things", "what's notable", "what stands out", "explore X", "what should I know about", "any findings", "anything unusual", "key insights"): ALWAYS call search_inside_documents with multiple diverse queries covering different aspects of the documents before answering. The document index is never sufficient for investigative answers — you must retrieve actual content.
              *     7. Never search for meta-queries like "summary", "overview", or "all documents" — search for the actual subject matter.]]
+             *     [[if view_figure: 8. Figures: a passage that is a figure (a chart, photo, diagram or signature) holds a GENERATED IMAGE CAPTION ("> [!ai-caption]") that a vision model wrote at upload so the figure can be found. It is not the document's text and not evidence. To answer about what an image shows, look at it with view_figure (its passage reference and the question) and say in your answer that you looked at the image with a vision model.]]
              *
              *     FORMATTING:
              *     - Start your answer directly with the content.
@@ -9924,7 +9925,7 @@ interface components {
              *       "AGENT_MAX_RUN_TOKENS": 1500000,
              *       "AGENT_MAX_SUBAGENT_SPAWNS": 4,
              *       "AGENT_HISTORY_TOKEN_BUDGET": 48000,
-             *       "AGENT_SYSTEM_PROMPT": "Answer any question the user asks. You have access to workspace documents — use them when relevant. When no documents are relevant, answer from your own knowledge. Always provide a helpful answer.\n\nWhen using documents, retrieve actual document content — do not rely on metadata or headings alone. Cross-check information across sources. Not all assertions in sources are factual — some are arguments by a party. Attribute claims to their source where parties disagree.\n\nDOCUMENT INDEX:\nThe \"Workspace Documents\" table in your system prompt lists every document with its Id, Title, Date, and Pages. The Subject field is a brief navigation label only — it does not contain enough information to answer questions about a document's content. Use this table directly when the user asks to list, catalogue, or tabulate documents — do NOT call any tools for that. Only use tools when you need to read or search document content.\n\nREADING TOOL RESULTS:\nEvery retrieval result opens with a CAPTION line naming the call that produced it, what it contains, and what it does not (for example \"FULL CONTENT … complete verbatim text, all 10 pages\" or \"DIGEST … verbatim page text is NOT part of this result\"). The caption also tells you when you already retrieved that material earlier in this turn. Read it before choosing your next call — it is the record of what you already hold.\n\nDOCUMENT STRATEGY, in order:\n1. Listing or cataloguing documents: use the document index already in your prompt — no tool calls needed.\n[[if get_table_of_contents, search_inside_documents, read_document: 2. Getting oriented in a document you have not read: call get_table_of_contents FIRST. It is cheap, returns the heading structure with page references, and tells you which pages actually matter. Do this before any read_document on a document longer than a few pages, rather than reading blind and then discovering the structure.\n3. Specific questions across several documents: search_inside_documents to locate the passages, then read_document on the pages it points to.\n4. Reading a document: read_document with page ranges for the sections you need, or with no ranges to load the complete text. A read with no ranges returns the WHOLE document — once you have it there is nothing further to retrieve from that document. Do not read it again, do not re-read ranges out of it, and do not ask for its outline afterwards.\n5. Prefer digest=true from the outset. It is the normal retrieval mode: findings with citations, compact context. Decide against it only when the task needs the exact words — translation, quotation, close reading of style or drafting. Choosing digest=false IS that decision, and it is final for that document: digest is a way of retrieving, not a way of processing text you already hold, so once you have loaded a document in full a digest of it is refused. Reason over the text you have.\n6. Open-ended investigative queries (\"find interesting things\", \"what's notable\", \"what stands out\", \"explore X\", \"what should I know about\", \"any findings\", \"anything unusual\", \"key insights\"): ALWAYS call search_inside_documents with multiple diverse queries covering different aspects of the documents before answering. The document index is never sufficient for investigative answers — you must retrieve actual content.\n7. Never search for meta-queries like \"summary\", \"overview\", or \"all documents\" — search for the actual subject matter.]]\n\nFORMATTING:\n- Start your answer directly with the content.\n- Use markdown: headings, bold, bullet points, numbered lists\n- Main conclusion first, followed by supporting points\n- Bullet points for lists of discrete items\n- Numbering for sequential steps or prioritized items\n- Bold text for critical conclusions or decision points\n\nTOOL CALLING:\n- Your job is to gather the context the question needs, then answer. Do not give a final answer while there is relevant context you have not yet retrieved — but material already in your context counts as retrieved, and fetching it a second time is not progress. If the question touches a workspace document, always check the source — never answer from a past turn's reply alone.\n- Always invoke tools using the tool-calling mechanism. Say what a call is doing in its `focus` argument.\n\nEFFICIENCY:\n1. Before each step, review the results of prior steps. Adapt your approach based on what you learned — if a call was rejected or returned poor results, adjust your strategy.\n2. Call multiple tools in parallel when a query has several aspects.\n3. NEVER repeat a tool call with the same arguments, and never re-retrieve material you already hold — the caption on each result tells you what that is.\n4. Use targeted queries and narrow page ranges per retrieval call.\n5. Stop when you have sufficient evidence — do not over-research."
+             *       "AGENT_SYSTEM_PROMPT": "Answer any question the user asks. You have access to workspace documents — use them when relevant. When no documents are relevant, answer from your own knowledge. Always provide a helpful answer.\n\nWhen using documents, retrieve actual document content — do not rely on metadata or headings alone. Cross-check information across sources. Not all assertions in sources are factual — some are arguments by a party. Attribute claims to their source where parties disagree.\n\nDOCUMENT INDEX:\nThe \"Workspace Documents\" table in your system prompt lists every document with its Id, Title, Date, and Pages. The Subject field is a brief navigation label only — it does not contain enough information to answer questions about a document's content. Use this table directly when the user asks to list, catalogue, or tabulate documents — do NOT call any tools for that. Only use tools when you need to read or search document content.\n\nREADING TOOL RESULTS:\nEvery retrieval result opens with a CAPTION line naming the call that produced it, what it contains, and what it does not (for example \"FULL CONTENT … complete verbatim text, all 10 pages\" or \"DIGEST … verbatim page text is NOT part of this result\"). The caption also tells you when you already retrieved that material earlier in this turn. Read it before choosing your next call — it is the record of what you already hold.\n\nDOCUMENT STRATEGY, in order:\n1. Listing or cataloguing documents: use the document index already in your prompt — no tool calls needed.\n[[if get_table_of_contents, search_inside_documents, read_document: 2. Getting oriented in a document you have not read: call get_table_of_contents FIRST. It is cheap, returns the heading structure with page references, and tells you which pages actually matter. Do this before any read_document on a document longer than a few pages, rather than reading blind and then discovering the structure.\n3. Specific questions across several documents: search_inside_documents to locate the passages, then read_document on the pages it points to.\n4. Reading a document: read_document with page ranges for the sections you need, or with no ranges to load the complete text. A read with no ranges returns the WHOLE document — once you have it there is nothing further to retrieve from that document. Do not read it again, do not re-read ranges out of it, and do not ask for its outline afterwards.\n5. Prefer digest=true from the outset. It is the normal retrieval mode: findings with citations, compact context. Decide against it only when the task needs the exact words — translation, quotation, close reading of style or drafting. Choosing digest=false IS that decision, and it is final for that document: digest is a way of retrieving, not a way of processing text you already hold, so once you have loaded a document in full a digest of it is refused. Reason over the text you have.\n6. Open-ended investigative queries (\"find interesting things\", \"what's notable\", \"what stands out\", \"explore X\", \"what should I know about\", \"any findings\", \"anything unusual\", \"key insights\"): ALWAYS call search_inside_documents with multiple diverse queries covering different aspects of the documents before answering. The document index is never sufficient for investigative answers — you must retrieve actual content.\n7. Never search for meta-queries like \"summary\", \"overview\", or \"all documents\" — search for the actual subject matter.]]\n[[if view_figure: 8. Figures: a passage that is a figure (a chart, photo, diagram or signature) holds a GENERATED IMAGE CAPTION (\"> [!ai-caption]\") that a vision model wrote at upload so the figure can be found. It is not the document's text and not evidence. To answer about what an image shows, look at it with view_figure (its passage reference and the question) and say in your answer that you looked at the image with a vision model.]]\n\nFORMATTING:\n- Start your answer directly with the content.\n- Use markdown: headings, bold, bullet points, numbered lists\n- Main conclusion first, followed by supporting points\n- Bullet points for lists of discrete items\n- Numbering for sequential steps or prioritized items\n- Bold text for critical conclusions or decision points\n\nTOOL CALLING:\n- Your job is to gather the context the question needs, then answer. Do not give a final answer while there is relevant context you have not yet retrieved — but material already in your context counts as retrieved, and fetching it a second time is not progress. If the question touches a workspace document, always check the source — never answer from a past turn's reply alone.\n- Always invoke tools using the tool-calling mechanism. Say what a call is doing in its `focus` argument.\n\nEFFICIENCY:\n1. Before each step, review the results of prior steps. Adapt your approach based on what you learned — if a call was rejected or returned poor results, adjust your strategy.\n2. Call multiple tools in parallel when a query has several aspects.\n3. NEVER repeat a tool call with the same arguments, and never re-retrieve material you already hold — the caption on each result tells you what that is.\n4. Use targeted queries and narrow page ranges per retrieval call.\n5. Stop when you have sufficient evidence — do not over-research."
              *     }
              */
             Agents: components['schemas']['AgentsConfig'];
@@ -9932,7 +9933,7 @@ interface components {
              * @default {
              *       "API_TYPE": "remote",
              *       "ENABLE_THINKING": false,
-             *       "MODEL_NAME": "dummy",
+             *       "MODEL_NAME": "Fast",
              *       "SYSTEM_INSTRUCTION": "You are ARBI, an AI assistant created by ARBI CITY. Be proactive, helpful and professional.Be concise. Use markdown where it helps readability. Cross-check information across sources before answering. Never fabricate facts. Not all assertions in the sources are factual — some are arguments by a party. Attribute claims to their source, especially where parties disagree.",
              *       "MAX_TOKEN_SIZE_TO_ANSWER": 200000,
              *       "TEMPERATURE": 0.1,
@@ -9944,7 +9945,7 @@ interface components {
              * @default {
              *       "API_TYPE": "remote",
              *       "ENABLE_THINKING": false,
-             *       "MODEL_NAME": "dummy",
+             *       "MODEL_NAME": "Wise",
              *       "SYSTEM_INSTRUCTION": "You are reviewing a draft answer prepared by a research agent. The draft was written based on summaries, but you now have access to the full source material.\n\nYour task:\n1. Review the draft answer against the source material provided\n2. Verify claims are supported by the sources\n3. Correct any inaccuracies or unsupported statements\n4. Add any important details from the sources that were missed\n5. Maintain formal, objective tone appropriate for professional contexts\n\nIf information is not found in the sources but the draft answer is based on web search results or general knowledge, preserve it and ensure it is clearly labelled as such. Only flag missing source support for claims that purport to come from the documents.\nDo not add inline citation markers - the system handles citations automatically.",
              *       "TEMPERATURE": 0.1,
              *       "MAX_TOKENS": 0,
@@ -9956,7 +9957,7 @@ interface components {
              * @default {
              *       "API_TYPE": "remote",
              *       "ENABLE_THINKING": false,
-             *       "MODEL_NAME": "dummy",
+             *       "MODEL_NAME": "Fast",
              *       "SYSTEM_INSTRUCTION": "You are a chunk evaluator. Analyze retrieved chunks to determine which are relevant to the user's query and extract key learnings from the RELEVANT chunks only. You must return your response as valid JSON matching the provided schema.",
              *       "TEMPERATURE": 0.1,
              *       "MAX_TOKENS": 0,
@@ -9971,7 +9972,7 @@ interface components {
              * @default {
              *       "API_TYPE": "remote",
              *       "ENABLE_THINKING": false,
-             *       "MODEL_NAME": "dummy",
+             *       "MODEL_NAME": "Fast",
              *       "SYSTEM_INSTRUCTION": "You are a title generator that creates concise, descriptive titles.\n\nOutput the title as plain text only. Do not use Markdown or any other formatting. Specifically:\n- Do not bold, italicize, or otherwise emphasize any part of the title.\n- Do not include asterisks (*), underscores (_), backticks (`), or other Markdown markers.\n- Do not prefix the title with bullets, dashes (-), or numbered/lettered enumeration (e.g. \"1.\", \"a)\").\n- Do not wrap the title in quotes or brackets.",
              *       "MAX_TOKEN_SIZE_TO_ANSWER": 20,
              *       "TEMPERATURE": 0.3,
@@ -9983,7 +9984,7 @@ interface components {
              * @default {
              *       "API_TYPE": "remote",
              *       "ENABLE_THINKING": false,
-             *       "MODEL_NAME": "dummy",
+             *       "MODEL_NAME": "Wise",
              *       "SYSTEM_INSTRUCTION": "You are a conversation summariser. Condense the conversation history into a concise summary that preserves:\n- Key decisions and conclusions reached\n- Specific names, dates, numbers, and document references\n- Open questions and unresolved items\n- The user's goals and constraints\n\nWrite in past tense, third person. Be specific, not vague. Do not omit important details in favour of brevity.",
              *       "TEMPERATURE": 0.1,
              *       "MAX_TOKENS": 0,
@@ -9997,7 +9998,7 @@ interface components {
              * @default {
              *       "API_TYPE": "remote",
              *       "ENABLE_THINKING": false,
-             *       "MODEL_NAME": "dummy",
+             *       "MODEL_NAME": "Fast",
              *       "SYSTEM_INSTRUCTION": "You are a document analysis assistant. Answer questions based on the document content.\n\nYour answers are displayed directly as cell values in a table, so keep every answer as short as possible.\n\nGuidelines:\n- Answer with the value only — never a full sentence. Do not restate the question or add lead-ins such as \"This document is...\" or \"The document describes...\". Good: \"German\". Bad: \"This document is written in German.\"\n- When an answer has multiple values, put each value on its own line: separate values with a newline character. Do not separate them with commas, bullets, dashes, numbering, or by writing them as a sentence.\n- Answer from the document. Reasonable inferences are fine; if the document gives no basis for an answer, the value is null and sources is empty.\n- Follow the format instructions exactly for each answer.\n- Output answers as plain text only. Do not use Markdown or any other formatting: no bold, italics, asterisks (*), underscores (_), backticks (`), bullets, or numbered/lettered enumeration.\n- Passages may be preceded by markers like [p3.2]: not document content. Cite them in \"sources\", never in a value.",
              *       "MAX_TOKEN_CONTEXT_TO_ANSWER": 200000,
              *       "METADATA_CONTEXT_TOKENS": 12000,
@@ -10059,7 +10060,7 @@ interface components {
              * @default {
              *       "API_TYPE": "remote",
              *       "ENABLE_THINKING": false,
-             *       "MODEL_NAME": "dummy",
+             *       "MODEL_NAME": "Fast",
              *       "SYSTEM_INSTRUCTION": "You are a professional legal translator. Translate each numbered passage into the target language, faithfully and completely: never summarise, shorten, merge, reorder or comment. Keep Markdown structure, numbers, dates, amounts, names, citations and cross-references exactly as written. Use the glossary translation for every glossary term, every time. Context passages are for reading only: never translate or return them.",
              *       "TEMPERATURE": 0,
              *       "MAX_TOKENS": 0,
@@ -10074,7 +10075,7 @@ interface components {
              * @default {
              *       "API_TYPE": "remote",
              *       "ENABLE_THINKING": false,
-             *       "MODEL_NAME": "dummy",
+             *       "MODEL_NAME": "Wise",
              *       "TEMPERATURE": 0.3,
              *       "MAX_TOKENS": 0,
              *       "MAX_TOKEN_CONTEXT": 200000,
@@ -10086,7 +10087,7 @@ interface components {
              * @default {
              *       "API_TYPE": "remote",
              *       "ENABLE_THINKING": true,
-             *       "MODEL_NAME": "dummy",
+             *       "MODEL_NAME": "Wise",
              *       "SYSTEM_INSTRUCTION": "You are a research planning assistant. Your job is to analyze a user query and produce a concise, numbered research plan.\n\nYou have access to a workspace of documents.[[if search_inside_documents, read_document, get_table_of_contents:  The available tools for executing the plan are:\n- search_inside_documents: Search workspace documents for relevant passages (semantic, keyword, or hybrid). Use digest=true to extract learnings.\n- read_document: Read page ranges or entire documents. Use digest=true to extract learnings with citations.\n- get_table_of_contents: Get document headings with page references]]\n\nConsider:\n- What information is needed to answer the query\n- Which documents are likely relevant based on the document index\n- What search queries and document passages to examine\n- What order of operations will be most efficient\n- Whether parallel searches can be used for different aspects\n\nRespond with ONLY the plan (numbered steps). Do not execute any steps.",
              *       "TEMPERATURE": 0.3,
              *       "MAX_TOKENS": 0,
@@ -10099,7 +10100,7 @@ interface components {
              * @default {
              *       "API_TYPE": "remote",
              *       "ENABLE_THINKING": false,
-             *       "MODEL_NAME": "dummy",
+             *       "MODEL_NAME": "Fast",
              *       "SYSTEM_INSTRUCTION": "You are a context selector. Given a user's query, conversation history, and a table of available documents, determine the user's goal and select relevant context.\n\n## Instructions\n\n1. **goal**: 1-2 concise sentences describing the user's intent AND your filtering rationale. Explain what documents are needed and why. Examples:\n   - \"The user is greeting the assistant. No documents are needed.\"\n   - \"The user wants to find force majeure clauses. Only the three contracts are relevant, so I will keep only those.\"\n   - \"The user wants a full timeline of the case. Most documents are relevant, so I will drop only the unrelated procedural orders.\"\n   - The agent that reads the goal does not see the history indices: refer to an earlier message by what it is (\"the investigation report in the previous answer\"), never by its [N] index.\n\n2. **doc_filter_mode** and **doc_ids**: Select which documents the agent needs.\n   - `\"keep\"` mode: `doc_ids` lists the ONLY docs to keep. Everything else is dropped.\n   - `\"drop\"` mode: `doc_ids` lists the ONLY docs to drop. Everything else is kept.\n   - Pick whichever mode produces the **shorter `doc_ids` list**.\n   - Doc IDs are in the **Id** column of the Available Documents table (e.g. `\"doc-abc12345\"`). Copy them **exactly** from the table — do not invent, guess, or modify IDs.\n\n   | Scenario | doc_filter_mode | doc_ids |\n   |----------|----------------|---------|\n   | Need ALL docs | `\"drop\"` | `[]` |\n   | Need NO docs | `\"keep\"` | `[]` |\n   | Need 3 of 100 docs | `\"keep\"` | `[\"id1\",\"id2\",\"id3\"]` |\n   | Drop 3 of 100 docs | `\"drop\"` | `[\"id1\",\"id2\",\"id3\"]` |\n\n3. **chat_range_start** and **chat_range_end**: 0-based inclusive range of earlier history messages that provide relevant context. The last 3 messages are always included automatically. Set both to -1 if no earlier history is needed.\n\n## Rules\n\n- **Never repeat a doc ID.** Each ID must appear at most once.\n- Before writing `doc_ids`, count how many you would keep vs drop, then pick the shorter list.\n- Greetings or meta-questions about the assistant (e.g. \"hello\", \"who are you?\", \"what can you do?\"): the agent does NOT need any documents → `\"keep\"` with `[]`.\n- Ambiguous references like \"the document\" or \"this contract\" without specifying which: keep all → `\"drop\"` with `[]`.\n- For follow-up questions, include history messages that establish the topic.\n- Always return valid JSON matching the schema.",
              *       "TEMPERATURE": 0,
              *       "MAX_TOKENS": 0
@@ -10110,7 +10111,7 @@ interface components {
              * @default {
              *       "API_TYPE": "remote",
              *       "ENABLE_THINKING": false,
-             *       "MODEL_NAME": "dummy",
+             *       "MODEL_NAME": "Fast",
              *       "TEMPERATURE": 0.1,
              *       "MAX_TOKENS": 0,
              *       "MAX_PAGES_PER_CALL": 5,
@@ -10131,7 +10132,7 @@ interface components {
              * @default {
              *       "API_TYPE": "remote",
              *       "ENABLE_THINKING": true,
-             *       "MODEL_NAME": "",
+             *       "MODEL_NAME": "Premium",
              *       "SYSTEM_INSTRUCTION": "You are a code execution agent. You write and run code to accomplish tasks.\n\nYou have access to an execute_code tool that runs code in a sandboxed Docker container.\n\nThe sandbox has:\n- Python 3.12 with numpy, pandas, matplotlib, requests, beautifulsoup4,\n  httpx, pillow, sympy, pyyaml (plus uv for installing more)\n- Node.js 22 with TypeScript (ts-node) and npm\n- Bash with git, curl, wget, jq, sqlite3, build-essential\n- Network access (can pip/npm install additional packages)\n\nInstructions:\n1. Write clean, correct code to accomplish the user's task\n2. Always call execute_code to run your code — never guess the output\n3. If execution errors, analyze the traceback, fix the code, and retry\n4. Once you have the correct output, respond with ONLY the final result\n\nKeep code simple and direct. Prefer Python unless the task specifically requires another language.",
              *       "TEMPERATURE": 0.2,
              *       "MAX_TOKENS": 0
@@ -10201,7 +10202,7 @@ interface components {
              *       },
              *       "MAX_CONCURRENT_REQUESTS": 256,
              *       "MODEL_NAME": "Qwen/Qwen3-Reranker-0.6B",
-             *       "API_TYPE": "local",
+             *       "API_TYPE": "remote",
              *       "RETRIEVAL_INSTRUCTION": "Judge whether the passage contains the specific facts, figures, names, or references mentioned in the query. Only answer yes if the passage directly contains these details"
              *     }
              */
@@ -10215,14 +10216,16 @@ interface components {
             /**
              * @default {
              *       "MAX_CHUNK_TOKENS": 1200,
-             *       "TOKENIZER_NAME": "cl100k_base"
+             *       "TOKENIZER_NAME": "cl100k_base",
+             *       "FIGURE_MIN_AREA": 0.005,
+             *       "FIGURE_MIN_SIDE": 0.02
              *     }
              */
             Chunker: components['schemas']['ChunkerConfig'];
             /**
              * @default {
              *       "MODEL_NAME": "Qwen/Qwen3-Embedding-0.6B",
-             *       "API_TYPE": "local",
+             *       "API_TYPE": "remote",
              *       "BATCH_SIZE": 128,
              *       "MAX_CONCURRENT_REQUESTS": 256,
              *       "DOCUMENT_PREFIX": "",
@@ -12550,6 +12553,12 @@ interface components {
             element_type?: string | null;
             /** Heading Level */
             heading_level?: number | null;
+            /**
+             * Chunk Kind
+             * @default text
+             * @enum {string}
+             */
+            chunk_kind: 'text' | 'figure';
         };
         /** ChunkerConfig */
         ChunkerConfig: {
@@ -12565,6 +12574,18 @@ interface components {
              * @default cl100k_base
              */
             TOKENIZER_NAME: string;
+            /**
+             * Figure Min Area
+             * @description A picture covering less than this fraction of its page (icons, bullets, small logos) is not made a figure chunk; it is logged instead.
+             * @default 0.005
+             */
+            FIGURE_MIN_AREA: number;
+            /**
+             * Figure Min Side
+             * @description A picture narrower or shorter than this fraction of its page (rules, lines) is not made a figure chunk; it is logged instead.
+             * @default 0.02
+             */
+            FIGURE_MIN_SIDE: number;
         };
         /**
          * CitationData
@@ -12660,7 +12681,7 @@ interface components {
             /**
              * Model Name
              * @description Model for code generation. Should be a strong coding model.
-             * @default
+             * @default Premium
              */
             MODEL_NAME: string;
             /**
@@ -15164,7 +15185,7 @@ interface components {
             /**
              * Model Name
              * @description The name of the non-reasoning model to be used.
-             * @default dummy
+             * @default Fast
              */
             MODEL_NAME: string;
             /**
@@ -15508,7 +15529,7 @@ interface components {
             /**
              * Api Type
              * @description The inference type (local or remote).
-             * @default local
+             * @default remote
              * @enum {string}
              */
             API_TYPE: 'local' | 'remote';
@@ -15631,7 +15652,7 @@ interface components {
             /**
              * Model Name
              * @description The name of the non-reasoning model to be used.
-             * @default dummy
+             * @default Fast
              */
             MODEL_NAME: string;
             /**
@@ -16236,7 +16257,7 @@ interface components {
             /**
              * Model Name
              * @description The name of the non-reasoning model to be used.
-             * @default dummy
+             * @default Fast
              */
             MODEL_NAME: string;
             /**
@@ -18090,7 +18111,7 @@ interface components {
             /**
              * Model Name
              * @description The model for note consolidation. Defaults to reasoning model.
-             * @default dummy
+             * @default Wise
              */
             MODEL_NAME: string;
             /**
@@ -18981,6 +19002,12 @@ interface components {
             bbox: number[];
             /** Text */
             text: string;
+            /**
+             * Kind
+             * @default text
+             * @enum {string}
+             */
+            kind: 'text' | 'figure';
         };
         /**
          * PageGeometry
@@ -19316,7 +19343,7 @@ interface components {
             /**
              * Model Name
              * @description The model for plan generation. Defaults to reasoning model.
-             * @default dummy
+             * @default Wise
              */
             MODEL_NAME: string;
             /**
@@ -20368,7 +20395,7 @@ interface components {
             /**
              * Model Name
              * @description The model for query answering.
-             * @default dummy
+             * @default Fast
              */
             MODEL_NAME: string;
             /**
@@ -20797,7 +20824,7 @@ interface components {
             /**
              * Api Type
              * @description The inference type (local or remote).
-             * @default local
+             * @default remote
              * @enum {string}
              */
             API_TYPE: 'local' | 'remote';
@@ -21657,7 +21684,7 @@ interface components {
             /**
              * Model Name
              * @description The model for reviewing agent draft answers. Defaults to reasoning model.
-             * @default dummy
+             * @default Wise
              */
             MODEL_NAME: string;
             /**
@@ -21869,6 +21896,7 @@ interface components {
             detail_read_document?: components['schemas']['ReadDocumentDetail'] | null;
             detail_get_table_of_contents?: components['schemas']['GetTableOfContentsDetail'] | null;
             detail_view_document_pages?: components['schemas']['ViewDocumentPagesDetail'] | null;
+            detail_view_figure?: components['schemas']['ViewFigureDetail'] | null;
             detail_web_search?: components['schemas']['WebSearchDetail'] | null;
             detail_personal_agent?: components['schemas']['PersonalAgentDetail'] | null;
             detail_read_url?: components['schemas']['ReadUrlDetail'] | null;
@@ -23358,7 +23386,7 @@ interface components {
             /**
              * Model Name
              * @description The model for conversation summarisation. Defaults to reasoning model.
-             * @default dummy
+             * @default Wise
              */
             MODEL_NAME: string;
             /**
@@ -24059,7 +24087,7 @@ interface components {
             /**
              * Model Name
              * @description The name of the non-reasoning model to be used.
-             * @default dummy
+             * @default Fast
              */
             MODEL_NAME: string;
             /**
@@ -24494,7 +24522,7 @@ interface components {
             /**
              * Model Name
              * @description The name of the non-reasoning model to be used.
-             * @default dummy
+             * @default Fast
              */
             MODEL_NAME: string;
             /**
@@ -25433,6 +25461,33 @@ interface components {
              */
             to_page: number;
         };
+        /**
+         * ViewFigureDetail
+         * @description Detail for a view_figure tool call (a vision model looks at a figure to answer a question).
+         */
+        ViewFigureDetail: {
+            /**
+             * Tool
+             * @default view_figure
+             * @constant
+             */
+            tool: 'view_figure';
+            /**
+             * Doc Ext Id
+             * @description Document external ID
+             */
+            doc_ext_id: string;
+            /**
+             * Figure
+             * @description The figure's passage reference ("4.2"), when named
+             */
+            figure?: string | null;
+            /**
+             * Page
+             * @description The page looked at
+             */
+            page?: number | null;
+        };
         /**
          * VisionLLMConfig
          * @description Configuration for VisionLLM — visually inspects document pages.
@@ -25458,7 +25513,7 @@ interface components {
             /**
              * Model Name
              * @description The vision-capable model name (must support image inputs).
-             * @default dummy
+             * @default Fast
              */
             MODEL_NAME: string;
             /**

package/dist/index.d.ts

@@ -9694,6 +9694,7 @@ interface components {
              *     5. Prefer digest=true from the outset. It is the normal retrieval mode: findings with citations, compact context. Decide against it only when the task needs the exact words — translation, quotation, close reading of style or drafting. Choosing digest=false IS that decision, and it is final for that document: digest is a way of retrieving, not a way of processing text you already hold, so once you have loaded a document in full a digest of it is refused. Reason over the text you have.
              *     6. Open-ended investigative queries ("find interesting things", "what's notable", "what stands out", "explore X", "what should I know about", "any findings", "anything unusual", "key insights"): ALWAYS call search_inside_documents with multiple diverse queries covering different aspects of the documents before answering. The document index is never sufficient for investigative answers — you must retrieve actual content.
              *     7. Never search for meta-queries like "summary", "overview", or "all documents" — search for the actual subject matter.]]
+             *     [[if view_figure: 8. Figures: a passage that is a figure (a chart, photo, diagram or signature) holds a GENERATED IMAGE CAPTION ("> [!ai-caption]") that a vision model wrote at upload so the figure can be found. It is not the document's text and not evidence. To answer about what an image shows, look at it with view_figure (its passage reference and the question) and say in your answer that you looked at the image with a vision model.]]
              *
              *     FORMATTING:
              *     - Start your answer directly with the content.
@@ -9924,7 +9925,7 @@ interface components {
              *       "AGENT_MAX_RUN_TOKENS": 1500000,
              *       "AGENT_MAX_SUBAGENT_SPAWNS": 4,
              *       "AGENT_HISTORY_TOKEN_BUDGET": 48000,
-             *       "AGENT_SYSTEM_PROMPT": "Answer any question the user asks. You have access to workspace documents — use them when relevant. When no documents are relevant, answer from your own knowledge. Always provide a helpful answer.\n\nWhen using documents, retrieve actual document content — do not rely on metadata or headings alone. Cross-check information across sources. Not all assertions in sources are factual — some are arguments by a party. Attribute claims to their source where parties disagree.\n\nDOCUMENT INDEX:\nThe \"Workspace Documents\" table in your system prompt lists every document with its Id, Title, Date, and Pages. The Subject field is a brief navigation label only — it does not contain enough information to answer questions about a document's content. Use this table directly when the user asks to list, catalogue, or tabulate documents — do NOT call any tools for that. Only use tools when you need to read or search document content.\n\nREADING TOOL RESULTS:\nEvery retrieval result opens with a CAPTION line naming the call that produced it, what it contains, and what it does not (for example \"FULL CONTENT … complete verbatim text, all 10 pages\" or \"DIGEST … verbatim page text is NOT part of this result\"). The caption also tells you when you already retrieved that material earlier in this turn. Read it before choosing your next call — it is the record of what you already hold.\n\nDOCUMENT STRATEGY, in order:\n1. Listing or cataloguing documents: use the document index already in your prompt — no tool calls needed.\n[[if get_table_of_contents, search_inside_documents, read_document: 2. Getting oriented in a document you have not read: call get_table_of_contents FIRST. It is cheap, returns the heading structure with page references, and tells you which pages actually matter. Do this before any read_document on a document longer than a few pages, rather than reading blind and then discovering the structure.\n3. Specific questions across several documents: search_inside_documents to locate the passages, then read_document on the pages it points to.\n4. Reading a document: read_document with page ranges for the sections you need, or with no ranges to load the complete text. A read with no ranges returns the WHOLE document — once you have it there is nothing further to retrieve from that document. Do not read it again, do not re-read ranges out of it, and do not ask for its outline afterwards.\n5. Prefer digest=true from the outset. It is the normal retrieval mode: findings with citations, compact context. Decide against it only when the task needs the exact words — translation, quotation, close reading of style or drafting. Choosing digest=false IS that decision, and it is final for that document: digest is a way of retrieving, not a way of processing text you already hold, so once you have loaded a document in full a digest of it is refused. Reason over the text you have.\n6. Open-ended investigative queries (\"find interesting things\", \"what's notable\", \"what stands out\", \"explore X\", \"what should I know about\", \"any findings\", \"anything unusual\", \"key insights\"): ALWAYS call search_inside_documents with multiple diverse queries covering different aspects of the documents before answering. The document index is never sufficient for investigative answers — you must retrieve actual content.\n7. Never search for meta-queries like \"summary\", \"overview\", or \"all documents\" — search for the actual subject matter.]]\n\nFORMATTING:\n- Start your answer directly with the content.\n- Use markdown: headings, bold, bullet points, numbered lists\n- Main conclusion first, followed by supporting points\n- Bullet points for lists of discrete items\n- Numbering for sequential steps or prioritized items\n- Bold text for critical conclusions or decision points\n\nTOOL CALLING:\n- Your job is to gather the context the question needs, then answer. Do not give a final answer while there is relevant context you have not yet retrieved — but material already in your context counts as retrieved, and fetching it a second time is not progress. If the question touches a workspace document, always check the source — never answer from a past turn's reply alone.\n- Always invoke tools using the tool-calling mechanism. Say what a call is doing in its `focus` argument.\n\nEFFICIENCY:\n1. Before each step, review the results of prior steps. Adapt your approach based on what you learned — if a call was rejected or returned poor results, adjust your strategy.\n2. Call multiple tools in parallel when a query has several aspects.\n3. NEVER repeat a tool call with the same arguments, and never re-retrieve material you already hold — the caption on each result tells you what that is.\n4. Use targeted queries and narrow page ranges per retrieval call.\n5. Stop when you have sufficient evidence — do not over-research."
+             *       "AGENT_SYSTEM_PROMPT": "Answer any question the user asks. You have access to workspace documents — use them when relevant. When no documents are relevant, answer from your own knowledge. Always provide a helpful answer.\n\nWhen using documents, retrieve actual document content — do not rely on metadata or headings alone. Cross-check information across sources. Not all assertions in sources are factual — some are arguments by a party. Attribute claims to their source where parties disagree.\n\nDOCUMENT INDEX:\nThe \"Workspace Documents\" table in your system prompt lists every document with its Id, Title, Date, and Pages. The Subject field is a brief navigation label only — it does not contain enough information to answer questions about a document's content. Use this table directly when the user asks to list, catalogue, or tabulate documents — do NOT call any tools for that. Only use tools when you need to read or search document content.\n\nREADING TOOL RESULTS:\nEvery retrieval result opens with a CAPTION line naming the call that produced it, what it contains, and what it does not (for example \"FULL CONTENT … complete verbatim text, all 10 pages\" or \"DIGEST … verbatim page text is NOT part of this result\"). The caption also tells you when you already retrieved that material earlier in this turn. Read it before choosing your next call — it is the record of what you already hold.\n\nDOCUMENT STRATEGY, in order:\n1. Listing or cataloguing documents: use the document index already in your prompt — no tool calls needed.\n[[if get_table_of_contents, search_inside_documents, read_document: 2. Getting oriented in a document you have not read: call get_table_of_contents FIRST. It is cheap, returns the heading structure with page references, and tells you which pages actually matter. Do this before any read_document on a document longer than a few pages, rather than reading blind and then discovering the structure.\n3. Specific questions across several documents: search_inside_documents to locate the passages, then read_document on the pages it points to.\n4. Reading a document: read_document with page ranges for the sections you need, or with no ranges to load the complete text. A read with no ranges returns the WHOLE document — once you have it there is nothing further to retrieve from that document. Do not read it again, do not re-read ranges out of it, and do not ask for its outline afterwards.\n5. Prefer digest=true from the outset. It is the normal retrieval mode: findings with citations, compact context. Decide against it only when the task needs the exact words — translation, quotation, close reading of style or drafting. Choosing digest=false IS that decision, and it is final for that document: digest is a way of retrieving, not a way of processing text you already hold, so once you have loaded a document in full a digest of it is refused. Reason over the text you have.\n6. Open-ended investigative queries (\"find interesting things\", \"what's notable\", \"what stands out\", \"explore X\", \"what should I know about\", \"any findings\", \"anything unusual\", \"key insights\"): ALWAYS call search_inside_documents with multiple diverse queries covering different aspects of the documents before answering. The document index is never sufficient for investigative answers — you must retrieve actual content.\n7. Never search for meta-queries like \"summary\", \"overview\", or \"all documents\" — search for the actual subject matter.]]\n[[if view_figure: 8. Figures: a passage that is a figure (a chart, photo, diagram or signature) holds a GENERATED IMAGE CAPTION (\"> [!ai-caption]\") that a vision model wrote at upload so the figure can be found. It is not the document's text and not evidence. To answer about what an image shows, look at it with view_figure (its passage reference and the question) and say in your answer that you looked at the image with a vision model.]]\n\nFORMATTING:\n- Start your answer directly with the content.\n- Use markdown: headings, bold, bullet points, numbered lists\n- Main conclusion first, followed by supporting points\n- Bullet points for lists of discrete items\n- Numbering for sequential steps or prioritized items\n- Bold text for critical conclusions or decision points\n\nTOOL CALLING:\n- Your job is to gather the context the question needs, then answer. Do not give a final answer while there is relevant context you have not yet retrieved — but material already in your context counts as retrieved, and fetching it a second time is not progress. If the question touches a workspace document, always check the source — never answer from a past turn's reply alone.\n- Always invoke tools using the tool-calling mechanism. Say what a call is doing in its `focus` argument.\n\nEFFICIENCY:\n1. Before each step, review the results of prior steps. Adapt your approach based on what you learned — if a call was rejected or returned poor results, adjust your strategy.\n2. Call multiple tools in parallel when a query has several aspects.\n3. NEVER repeat a tool call with the same arguments, and never re-retrieve material you already hold — the caption on each result tells you what that is.\n4. Use targeted queries and narrow page ranges per retrieval call.\n5. Stop when you have sufficient evidence — do not over-research."
              *     }
              */
             Agents: components['schemas']['AgentsConfig'];
@@ -9932,7 +9933,7 @@ interface components {
              * @default {
              *       "API_TYPE": "remote",
              *       "ENABLE_THINKING": false,
-             *       "MODEL_NAME": "dummy",
+             *       "MODEL_NAME": "Fast",
              *       "SYSTEM_INSTRUCTION": "You are ARBI, an AI assistant created by ARBI CITY. Be proactive, helpful and professional.Be concise. Use markdown where it helps readability. Cross-check information across sources before answering. Never fabricate facts. Not all assertions in the sources are factual — some are arguments by a party. Attribute claims to their source, especially where parties disagree.",
              *       "MAX_TOKEN_SIZE_TO_ANSWER": 200000,
              *       "TEMPERATURE": 0.1,
@@ -9944,7 +9945,7 @@ interface components {
              * @default {
              *       "API_TYPE": "remote",
              *       "ENABLE_THINKING": false,
-             *       "MODEL_NAME": "dummy",
+             *       "MODEL_NAME": "Wise",
              *       "SYSTEM_INSTRUCTION": "You are reviewing a draft answer prepared by a research agent. The draft was written based on summaries, but you now have access to the full source material.\n\nYour task:\n1. Review the draft answer against the source material provided\n2. Verify claims are supported by the sources\n3. Correct any inaccuracies or unsupported statements\n4. Add any important details from the sources that were missed\n5. Maintain formal, objective tone appropriate for professional contexts\n\nIf information is not found in the sources but the draft answer is based on web search results or general knowledge, preserve it and ensure it is clearly labelled as such. Only flag missing source support for claims that purport to come from the documents.\nDo not add inline citation markers - the system handles citations automatically.",
              *       "TEMPERATURE": 0.1,
              *       "MAX_TOKENS": 0,
@@ -9956,7 +9957,7 @@ interface components {
              * @default {
              *       "API_TYPE": "remote",
              *       "ENABLE_THINKING": false,
-             *       "MODEL_NAME": "dummy",
+             *       "MODEL_NAME": "Fast",
              *       "SYSTEM_INSTRUCTION": "You are a chunk evaluator. Analyze retrieved chunks to determine which are relevant to the user's query and extract key learnings from the RELEVANT chunks only. You must return your response as valid JSON matching the provided schema.",
              *       "TEMPERATURE": 0.1,
              *       "MAX_TOKENS": 0,
@@ -9971,7 +9972,7 @@ interface components {
              * @default {
              *       "API_TYPE": "remote",
              *       "ENABLE_THINKING": false,
-             *       "MODEL_NAME": "dummy",
+             *       "MODEL_NAME": "Fast",
              *       "SYSTEM_INSTRUCTION": "You are a title generator that creates concise, descriptive titles.\n\nOutput the title as plain text only. Do not use Markdown or any other formatting. Specifically:\n- Do not bold, italicize, or otherwise emphasize any part of the title.\n- Do not include asterisks (*), underscores (_), backticks (`), or other Markdown markers.\n- Do not prefix the title with bullets, dashes (-), or numbered/lettered enumeration (e.g. \"1.\", \"a)\").\n- Do not wrap the title in quotes or brackets.",
              *       "MAX_TOKEN_SIZE_TO_ANSWER": 20,
              *       "TEMPERATURE": 0.3,
@@ -9983,7 +9984,7 @@ interface components {
              * @default {
              *       "API_TYPE": "remote",
              *       "ENABLE_THINKING": false,
-             *       "MODEL_NAME": "dummy",
+             *       "MODEL_NAME": "Wise",
              *       "SYSTEM_INSTRUCTION": "You are a conversation summariser. Condense the conversation history into a concise summary that preserves:\n- Key decisions and conclusions reached\n- Specific names, dates, numbers, and document references\n- Open questions and unresolved items\n- The user's goals and constraints\n\nWrite in past tense, third person. Be specific, not vague. Do not omit important details in favour of brevity.",
              *       "TEMPERATURE": 0.1,
              *       "MAX_TOKENS": 0,
@@ -9997,7 +9998,7 @@ interface components {
              * @default {
              *       "API_TYPE": "remote",
              *       "ENABLE_THINKING": false,
-             *       "MODEL_NAME": "dummy",
+             *       "MODEL_NAME": "Fast",
              *       "SYSTEM_INSTRUCTION": "You are a document analysis assistant. Answer questions based on the document content.\n\nYour answers are displayed directly as cell values in a table, so keep every answer as short as possible.\n\nGuidelines:\n- Answer with the value only — never a full sentence. Do not restate the question or add lead-ins such as \"This document is...\" or \"The document describes...\". Good: \"German\". Bad: \"This document is written in German.\"\n- When an answer has multiple values, put each value on its own line: separate values with a newline character. Do not separate them with commas, bullets, dashes, numbering, or by writing them as a sentence.\n- Answer from the document. Reasonable inferences are fine; if the document gives no basis for an answer, the value is null and sources is empty.\n- Follow the format instructions exactly for each answer.\n- Output answers as plain text only. Do not use Markdown or any other formatting: no bold, italics, asterisks (*), underscores (_), backticks (`), bullets, or numbered/lettered enumeration.\n- Passages may be preceded by markers like [p3.2]: not document content. Cite them in \"sources\", never in a value.",
              *       "MAX_TOKEN_CONTEXT_TO_ANSWER": 200000,
              *       "METADATA_CONTEXT_TOKENS": 12000,
@@ -10059,7 +10060,7 @@ interface components {
              * @default {
              *       "API_TYPE": "remote",
              *       "ENABLE_THINKING": false,
-             *       "MODEL_NAME": "dummy",
+             *       "MODEL_NAME": "Fast",
              *       "SYSTEM_INSTRUCTION": "You are a professional legal translator. Translate each numbered passage into the target language, faithfully and completely: never summarise, shorten, merge, reorder or comment. Keep Markdown structure, numbers, dates, amounts, names, citations and cross-references exactly as written. Use the glossary translation for every glossary term, every time. Context passages are for reading only: never translate or return them.",
              *       "TEMPERATURE": 0,
              *       "MAX_TOKENS": 0,
@@ -10074,7 +10075,7 @@ interface components {
              * @default {
              *       "API_TYPE": "remote",
              *       "ENABLE_THINKING": false,
-             *       "MODEL_NAME": "dummy",
+             *       "MODEL_NAME": "Wise",
              *       "TEMPERATURE": 0.3,
              *       "MAX_TOKENS": 0,
              *       "MAX_TOKEN_CONTEXT": 200000,
@@ -10086,7 +10087,7 @@ interface components {
              * @default {
              *       "API_TYPE": "remote",
              *       "ENABLE_THINKING": true,
-             *       "MODEL_NAME": "dummy",
+             *       "MODEL_NAME": "Wise",
              *       "SYSTEM_INSTRUCTION": "You are a research planning assistant. Your job is to analyze a user query and produce a concise, numbered research plan.\n\nYou have access to a workspace of documents.[[if search_inside_documents, read_document, get_table_of_contents:  The available tools for executing the plan are:\n- search_inside_documents: Search workspace documents for relevant passages (semantic, keyword, or hybrid). Use digest=true to extract learnings.\n- read_document: Read page ranges or entire documents. Use digest=true to extract learnings with citations.\n- get_table_of_contents: Get document headings with page references]]\n\nConsider:\n- What information is needed to answer the query\n- Which documents are likely relevant based on the document index\n- What search queries and document passages to examine\n- What order of operations will be most efficient\n- Whether parallel searches can be used for different aspects\n\nRespond with ONLY the plan (numbered steps). Do not execute any steps.",
              *       "TEMPERATURE": 0.3,
              *       "MAX_TOKENS": 0,
@@ -10099,7 +10100,7 @@ interface components {
              * @default {
              *       "API_TYPE": "remote",
              *       "ENABLE_THINKING": false,
-             *       "MODEL_NAME": "dummy",
+             *       "MODEL_NAME": "Fast",
              *       "SYSTEM_INSTRUCTION": "You are a context selector. Given a user's query, conversation history, and a table of available documents, determine the user's goal and select relevant context.\n\n## Instructions\n\n1. **goal**: 1-2 concise sentences describing the user's intent AND your filtering rationale. Explain what documents are needed and why. Examples:\n   - \"The user is greeting the assistant. No documents are needed.\"\n   - \"The user wants to find force majeure clauses. Only the three contracts are relevant, so I will keep only those.\"\n   - \"The user wants a full timeline of the case. Most documents are relevant, so I will drop only the unrelated procedural orders.\"\n   - The agent that reads the goal does not see the history indices: refer to an earlier message by what it is (\"the investigation report in the previous answer\"), never by its [N] index.\n\n2. **doc_filter_mode** and **doc_ids**: Select which documents the agent needs.\n   - `\"keep\"` mode: `doc_ids` lists the ONLY docs to keep. Everything else is dropped.\n   - `\"drop\"` mode: `doc_ids` lists the ONLY docs to drop. Everything else is kept.\n   - Pick whichever mode produces the **shorter `doc_ids` list**.\n   - Doc IDs are in the **Id** column of the Available Documents table (e.g. `\"doc-abc12345\"`). Copy them **exactly** from the table — do not invent, guess, or modify IDs.\n\n   | Scenario | doc_filter_mode | doc_ids |\n   |----------|----------------|---------|\n   | Need ALL docs | `\"drop\"` | `[]` |\n   | Need NO docs | `\"keep\"` | `[]` |\n   | Need 3 of 100 docs | `\"keep\"` | `[\"id1\",\"id2\",\"id3\"]` |\n   | Drop 3 of 100 docs | `\"drop\"` | `[\"id1\",\"id2\",\"id3\"]` |\n\n3. **chat_range_start** and **chat_range_end**: 0-based inclusive range of earlier history messages that provide relevant context. The last 3 messages are always included automatically. Set both to -1 if no earlier history is needed.\n\n## Rules\n\n- **Never repeat a doc ID.** Each ID must appear at most once.\n- Before writing `doc_ids`, count how many you would keep vs drop, then pick the shorter list.\n- Greetings or meta-questions about the assistant (e.g. \"hello\", \"who are you?\", \"what can you do?\"): the agent does NOT need any documents → `\"keep\"` with `[]`.\n- Ambiguous references like \"the document\" or \"this contract\" without specifying which: keep all → `\"drop\"` with `[]`.\n- For follow-up questions, include history messages that establish the topic.\n- Always return valid JSON matching the schema.",
              *       "TEMPERATURE": 0,
              *       "MAX_TOKENS": 0
@@ -10110,7 +10111,7 @@ interface components {
              * @default {
              *       "API_TYPE": "remote",
              *       "ENABLE_THINKING": false,
-             *       "MODEL_NAME": "dummy",
+             *       "MODEL_NAME": "Fast",
              *       "TEMPERATURE": 0.1,
              *       "MAX_TOKENS": 0,
              *       "MAX_PAGES_PER_CALL": 5,
@@ -10131,7 +10132,7 @@ interface components {
              * @default {
              *       "API_TYPE": "remote",
              *       "ENABLE_THINKING": true,
-             *       "MODEL_NAME": "",
+             *       "MODEL_NAME": "Premium",
              *       "SYSTEM_INSTRUCTION": "You are a code execution agent. You write and run code to accomplish tasks.\n\nYou have access to an execute_code tool that runs code in a sandboxed Docker container.\n\nThe sandbox has:\n- Python 3.12 with numpy, pandas, matplotlib, requests, beautifulsoup4,\n  httpx, pillow, sympy, pyyaml (plus uv for installing more)\n- Node.js 22 with TypeScript (ts-node) and npm\n- Bash with git, curl, wget, jq, sqlite3, build-essential\n- Network access (can pip/npm install additional packages)\n\nInstructions:\n1. Write clean, correct code to accomplish the user's task\n2. Always call execute_code to run your code — never guess the output\n3. If execution errors, analyze the traceback, fix the code, and retry\n4. Once you have the correct output, respond with ONLY the final result\n\nKeep code simple and direct. Prefer Python unless the task specifically requires another language.",
              *       "TEMPERATURE": 0.2,
              *       "MAX_TOKENS": 0
@@ -10201,7 +10202,7 @@ interface components {
              *       },
              *       "MAX_CONCURRENT_REQUESTS": 256,
              *       "MODEL_NAME": "Qwen/Qwen3-Reranker-0.6B",
-             *       "API_TYPE": "local",
+             *       "API_TYPE": "remote",
              *       "RETRIEVAL_INSTRUCTION": "Judge whether the passage contains the specific facts, figures, names, or references mentioned in the query. Only answer yes if the passage directly contains these details"
              *     }
              */
@@ -10215,14 +10216,16 @@ interface components {
             /**
              * @default {
              *       "MAX_CHUNK_TOKENS": 1200,
-             *       "TOKENIZER_NAME": "cl100k_base"
+             *       "TOKENIZER_NAME": "cl100k_base",
+             *       "FIGURE_MIN_AREA": 0.005,
+             *       "FIGURE_MIN_SIDE": 0.02
              *     }
              */
             Chunker: components['schemas']['ChunkerConfig'];
             /**
              * @default {
              *       "MODEL_NAME": "Qwen/Qwen3-Embedding-0.6B",
-             *       "API_TYPE": "local",
+             *       "API_TYPE": "remote",
              *       "BATCH_SIZE": 128,
              *       "MAX_CONCURRENT_REQUESTS": 256,
              *       "DOCUMENT_PREFIX": "",
@@ -12550,6 +12553,12 @@ interface components {
             element_type?: string | null;
             /** Heading Level */
             heading_level?: number | null;
+            /**
+             * Chunk Kind
+             * @default text
+             * @enum {string}
+             */
+            chunk_kind: 'text' | 'figure';
         };
         /** ChunkerConfig */
         ChunkerConfig: {
@@ -12565,6 +12574,18 @@ interface components {
              * @default cl100k_base
              */
             TOKENIZER_NAME: string;
+            /**
+             * Figure Min Area
+             * @description A picture covering less than this fraction of its page (icons, bullets, small logos) is not made a figure chunk; it is logged instead.
+             * @default 0.005
+             */
+            FIGURE_MIN_AREA: number;
+            /**
+             * Figure Min Side
+             * @description A picture narrower or shorter than this fraction of its page (rules, lines) is not made a figure chunk; it is logged instead.
+             * @default 0.02
+             */
+            FIGURE_MIN_SIDE: number;
         };
         /**
          * CitationData
@@ -12660,7 +12681,7 @@ interface components {
             /**
              * Model Name
              * @description Model for code generation. Should be a strong coding model.
-             * @default
+             * @default Premium
              */
             MODEL_NAME: string;
             /**
@@ -15164,7 +15185,7 @@ interface components {
             /**
              * Model Name
              * @description The name of the non-reasoning model to be used.
-             * @default dummy
+             * @default Fast
              */
             MODEL_NAME: string;
             /**
@@ -15508,7 +15529,7 @@ interface components {
             /**
              * Api Type
              * @description The inference type (local or remote).
-             * @default local
+             * @default remote
              * @enum {string}
              */
             API_TYPE: 'local' | 'remote';
@@ -15631,7 +15652,7 @@ interface components {
             /**
              * Model Name
              * @description The name of the non-reasoning model to be used.
-             * @default dummy
+             * @default Fast
              */
             MODEL_NAME: string;
             /**
@@ -16236,7 +16257,7 @@ interface components {
             /**
              * Model Name
              * @description The name of the non-reasoning model to be used.
-             * @default dummy
+             * @default Fast
              */
             MODEL_NAME: string;
             /**
@@ -18090,7 +18111,7 @@ interface components {
             /**
              * Model Name
              * @description The model for note consolidation. Defaults to reasoning model.
-             * @default dummy
+             * @default Wise
              */
             MODEL_NAME: string;
             /**
@@ -18981,6 +19002,12 @@ interface components {
             bbox: number[];
             /** Text */
             text: string;
+            /**
+             * Kind
+             * @default text
+             * @enum {string}
+             */
+            kind: 'text' | 'figure';
         };
         /**
          * PageGeometry
@@ -19316,7 +19343,7 @@ interface components {
             /**
              * Model Name
              * @description The model for plan generation. Defaults to reasoning model.
-             * @default dummy
+             * @default Wise
              */
             MODEL_NAME: string;
             /**
@@ -20368,7 +20395,7 @@ interface components {
             /**
              * Model Name
              * @description The model for query answering.
-             * @default dummy
+             * @default Fast
              */
             MODEL_NAME: string;
             /**
@@ -20797,7 +20824,7 @@ interface components {
             /**
              * Api Type
              * @description The inference type (local or remote).
-             * @default local
+             * @default remote
              * @enum {string}
              */
             API_TYPE: 'local' | 'remote';
@@ -21657,7 +21684,7 @@ interface components {
             /**
              * Model Name
              * @description The model for reviewing agent draft answers. Defaults to reasoning model.
-             * @default dummy
+             * @default Wise
              */
             MODEL_NAME: string;
             /**
@@ -21869,6 +21896,7 @@ interface components {
             detail_read_document?: components['schemas']['ReadDocumentDetail'] | null;
             detail_get_table_of_contents?: components['schemas']['GetTableOfContentsDetail'] | null;
             detail_view_document_pages?: components['schemas']['ViewDocumentPagesDetail'] | null;
+            detail_view_figure?: components['schemas']['ViewFigureDetail'] | null;
             detail_web_search?: components['schemas']['WebSearchDetail'] | null;
             detail_personal_agent?: components['schemas']['PersonalAgentDetail'] | null;
             detail_read_url?: components['schemas']['ReadUrlDetail'] | null;
@@ -23358,7 +23386,7 @@ interface components {
             /**
              * Model Name
              * @description The model for conversation summarisation. Defaults to reasoning model.
-             * @default dummy
+             * @default Wise
              */
             MODEL_NAME: string;
             /**
@@ -24059,7 +24087,7 @@ interface components {
             /**
              * Model Name
              * @description The name of the non-reasoning model to be used.
-             * @default dummy
+             * @default Fast
              */
             MODEL_NAME: string;
             /**
@@ -24494,7 +24522,7 @@ interface components {
             /**
              * Model Name
              * @description The name of the non-reasoning model to be used.
-             * @default dummy
+             * @default Fast
              */
             MODEL_NAME: string;
             /**
@@ -25433,6 +25461,33 @@ interface components {
              */
             to_page: number;
         };
+        /**
+         * ViewFigureDetail
+         * @description Detail for a view_figure tool call (a vision model looks at a figure to answer a question).
+         */
+        ViewFigureDetail: {
+            /**
+             * Tool
+             * @default view_figure
+             * @constant
+             */
+            tool: 'view_figure';
+            /**
+             * Doc Ext Id
+             * @description Document external ID
+             */
+            doc_ext_id: string;
+            /**
+             * Figure
+             * @description The figure's passage reference ("4.2"), when named
+             */
+            figure?: string | null;
+            /**
+             * Page
+             * @description The page looked at
+             */
+            page?: number | null;
+        };
         /**
          * VisionLLMConfig
          * @description Configuration for VisionLLM — visually inspects document pages.
@@ -25458,7 +25513,7 @@ interface components {
             /**
              * Model Name
              * @description The vision-capable model name (must support image inputs).
-             * @default dummy
+             * @default Fast
              */
             MODEL_NAME: string;
             /**

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@arbidocs/client",
-  "version": "0.5.122",
+  "version": "0.5.123",
   "description": "TypeScript SDK for the ARBI API — zero-knowledge auth, E2E encryption, and type-safe REST client",
   "type": "module",
   "sideEffects": false,