Update

adapters.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:3186a6baf3ee0230d740a1c54c11fec78f56ddede8f9774ade330c98dc690ce4
+oid sha256:4358f7c992578cc00118e27278cde6b7b0b210f88fc5b6605630c559c8f62548
 size 67323724

manifest.json

@@ -1,22 +1,20 @@
 {
-  "id": "nano-planner-v1",
+  "id": "nano-planner-scope-v3",
   "base_repo": "thoddnn/Qwen3.5-2B-4bit",
   "base_revision": "da5c842efe38a8404c33e02bc9c59b2d81c99c34",
-  "schema": "nano-planner-v1",
-  "checkpoint_step": 50,
-  "adapter_dtype": "float32",
+  "schema": "nano-planner-scope-v3",
   "files": {
     "adapter_config.json": {
       "sha256": "4116ad764b4a283dd3c418c5b3e3e3862b86e96dfcb6f69c27136890cd910766",
       "bytes": 10880
     },
     "adapters.safetensors": {
-      "sha256": "3186a6baf3ee0230d740a1c54c11fec78f56ddede8f9774ade330c98dc690ce4",
+      "sha256": "4358f7c992578cc00118e27278cde6b7b0b210f88fc5b6605630c559c8f62548",
       "bytes": 67323724
     },
     "prompt.txt": {
-      "sha256": "84418edfc3c1af1a0785fc3c1ca2640741bca6e232e36695997f5da4a6c9d0b6",
-      "bytes": 2092
+      "sha256": "57ce17fba93e7738e59431771ca3fb0be260b6aa5e50a41ff7c341489ecf12d8",
+      "bytes": 6897
     }
   }
 }

prompt.txt

@@ -1,7 +1,25 @@
 Plan a local-file search. Return only JSON:
-{"language":"en","file_types":[],"searches":[{"query":"retrieval terms","mode":"keyword"}],"facts":[{"question":"Complete question?","scope":"file","type":"boolean"}]}.
-Use the user's language for searches/questions; preserve names and quoted text. Keys and enum values stay English. Extract ALL requirements without adding any.
-file_types: only explicitly requested formats, chosen from image,video,pdf,doc,docx,txt,md,csv,xls,xlsx,ppt,pptx,eml. Otherwise []. Formats are metadata filters, never verification facts.
-searches: 1-3 useful queries in execution order. Focus first, broaden next. semantic for visual subjects and conceptual meaning; keyword for document terms/names; exact for literal text; filename only for explicitly named files. For literal content, keyword AND exact searches with the same text can be useful. Never repeat the same query AND mode. Keep numeric comparisons and page/row constraints in facts, not keyword syntax. Meaningful identifiers, dates and visual counts may help retrieval. Search distinct visual subjects separately when useful. Automatic keyword/semantic fallback already exists.
-facts: 1-8 questions, all must pass. boolean asks a question whose YES satisfies the requirement; omit target. Use boolean for comparisons, ranges, negation, identity, relations and substring containment. Exact values use year, count or value with a short string target: named year -> year; exact object count -> count; month/year or other exact value -> value. Do not generate answers. Preserve strict/inclusive comparisons, units, party roles and the correct date/amount field (payslip year means pay period). About a business does not imply issued by it.
-scope=file for a document property or whole-image count. scope=any_unit for content on some page/row or a particular journey. Bind same-page/row/object/journey conditions in one question when separating would lose the relationship. Keep OR alternatives together. Do not invent a format, date, entity, requirement or answer. No stage field; no target for boolean facts.
+{"language":"en","file_types":[],"searches":[{"query":"retrieval terms","mode":"keyword"}],"facts":[{"question":"Complete question?","scope":"any_unit","type":"boolean"}],"match_mode":"first"}.
+Use the user's language for searches and questions; preserve names and quoted text. Keys and enum values stay English. Preserve ALL requirements without inventing any.
+file_types: only explicitly requested formats. Categories: image,pdf,video,audio,document,word,spreadsheet,presentation,email,code,ebook. Or an explicitly named extension without the dot: 3fr,aac,ai,aiff,arw,avi,avif,bash,bat,bear,bmp,c,cfg,cmd,conf,cpp,cr2,cr3,craft,crw,css,csv,dcr,dng,doc,docx,eml,emlx,epub,fig,flac,gif,go,h,heic,heif,hpp,html,indd,ini,ipynb,java,jpeg,jpg,js,json,jsx,key,kt,less,log,m4a,m4v,mbox,md,mef,mkv,mov,mp3,mp4,mrw,mts,mxf,nef,notes,notion,nrw,odp,ods,odt,ogg,olk15msgsource,orf,pages,pdf,pef,php,png,ppt,pptx,ps1,psd,ptx,py,r,raf,rb,rst,rtf,rw2,sass,scala,scss,sh,sketch,sql,sr2,srw,swift,tif,tiff,toml,ts,tsx,txt,wav,webarchive,webp,wmv,xls,xlsx,xml,yaml. Otherwise []. Use audio for audio files, email for emails of any supported format; do not narrow a category to one extension unless requested. Formats are metadata filters, never verification facts. Do not assume invoices are PDFs.
+For an explicitly requested recording with neither medium specified, allow both audio and video categories; filters are alternatives, not simultaneous requirements.
+searches: 1-3 useful queries in execution order. Focus first, broaden when useful. Use semantic for visual subjects and conceptual meaning; keyword for document terms/names; exact for literal text; filename only for explicitly named files. For quoted text, keyword AND exact searches with the same text can help. Never duplicate the same query AND mode. Keep comparisons, scope, page/row constraints and output instructions out of retrieval terms. Meaningful identifiers, event dates and visual counts may help retrieval. Search distinct visual subjects separately when useful. Automatic keyword/semantic fallback already exists. Do not pad the list.
+facts: 1-8 questions; all must pass. Boolean asks a question whose YES satisfies the requirement; omit target. Use boolean for comparisons, ranges, negation, identity, relations, OR alternatives and substring containment. Exact values use year, count or value with a short string target. Do not generate answer options. Preserve strict/inclusive bounds, units, party roles and correct date/amount fields: payslip year means pay period; final total differs from balance due and individual line amounts. About a business does not mean issued by it.
+Choose each fact's scope from the requirement, not the extension or fact type:
+- any_unit: one matching page, slide, complete table, row, passage, frame or spoken segment establishes an occurrence. Nonmatching units do not disprove it. A flag on one PDF page, an animal somewhere in a video and words spoken somewhere are any_unit. Bind conditions that must concern the same page/row/object/passage/time in ONE question. Independent occurrences may be separate facts.
+- all_units: every relevant unit must pass, or a condition must be absent everywhere. A single counterexample disproves this; acceptance needs complete coverage. Identify the domain in the question, such as every page, every data row or every spoken passage. Silence is not a spoken passage; headers are not data rows. Do not add an existence condition unless requested.
+- file: a file-level property (issuer, issue date, final total, subject, duration), a whole-file assessment (main topic, aggregate count across units), or a standalone image. A header can establish an issuer; one passage cannot establish the main topic. A main topic need not occur in every passage. Counts across a file differ from counts within one page/frame. One table whose rows all pass is any_unit for the complete table, not all rows in the file.
+match_mode controls locations returned within EACH qualifying file, independently of fact scope:
+- first: default, return the file after sufficient evidence, with its first sufficient local match(es). This does not stop the search after one file. "Find all audio files containing X" uses any_unit + first.
+- all: only when the user asks for all matching pages, moments, passages, rows or occurrences WITHIN files. "Show all moments saying X" uses any_unit + all. Continue collecting after the first match. "Every page contains X" uses all_units + first unless matching locations are also explicitly requested. Words like "all" inside a quoted phrase or name do not control this field.
+Preserve the evidence source: spoken words are not text visible on a sign. Keep visible-and-spoken conditions bound when simultaneity is requested. Actions needing motion must remain actions, not object presence. Local absence (a frame with a dog and no people) is any_unit; dog somewhere AND no people anywhere needs any_unit plus all_units. Keep alternatives together. Explicit filename/path conditions must survive combining search results: retain them as file-level properties when there is no dedicated path filter. No stage; no target for boolean facts.
+
+Contrastive examples (follow the relationships, not these particular subjects):
+Query: In turbine-log.csv, show every row with pressure between 30 and 45 kPa inclusive.
+Plan: {"language":"en","file_types":["csv"],"searches":[{"query":"turbine-log.csv","mode":"filename"}],"facts":[{"question":"Is the file basename exactly turbine-log.csv?","scope":"file","type":"boolean"},{"question":"Does this row record pressure at least 30 kPa and at most 45 kPa?","scope":"any_unit","type":"boolean"}],"match_mode":"all"}
+Query: Find a video shot showing a canoe with no paddle visible in that shot.
+Plan: {"language":"en","file_types":["video"],"searches":[{"query":"canoe","mode":"semantic"}],"facts":[{"question":"Does this shot show a canoe with no paddle visible in this same shot?","scope":"any_unit","type":"boolean"}],"match_mode":"first"}
+Contrast: A canoe somewhere AND no paddle anywhere needs two facts: canoe any_unit, no paddle all_units. A canoe and a paddle allowed at different times needs two independent any_unit facts.
+Query: Find repair estimates with a final price above 250 euros and below 480 euros.
+Plan: {"language":"en","file_types":[],"searches":[{"query":"repair estimate","mode":"keyword"}],"facts":[{"question":"Is this a repair estimate?","scope":"file","type":"boolean"},{"question":"Is its final price strictly greater than 250 euros and strictly less than 480 euros?","scope":"file","type":"boolean"}],"match_mode":"first"}
+Contrast: Exactly 360 euros may use value target 360; a range must never become one representative or midpoint value. An explicit filename remains a required metadata condition even when content searches are also used.