seb/dashback
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

u

Browse Source
This commit is contained in:
seb
2026-10-06 03:30:31 +02:00
parent 0369b4e819
commit d5e58e76e1
15 changed files with 103 additions and 196 deletions
Download Patch File Download Diff File Expand all files Collapse all files

67
src/services/conversationManager.js
View File

@@ -105,52 +105,51 @@ async function buildSystemPrompt(vfsRoot, googleAuthId) {
}
const today = new Date().toISOString().split('T')[0];
const basePrompt = [
'You are an intelligent assistant specialized in exploring, querying, and managing data from the Dashback REST API.',
`Today's date: ${today}`,
'Assistant for QuixPOS Article Management System REST API operation',
`Date: ${today}`,
'',
'YOUR TOOLS:',
'1. dashback_request: Execute HTTP requests (GET, POST, PUT, PATCH, DELETE) against Dashback endpoints.',
'2. ask_user: Ask clarifying questions or request decisions from the user if their intent is ambiguous or requires choices. When you call ask_user, execution pauses immediately and waits for the user to provide their response.',
'3. javascript_execute: Run JavaScript snippets in a sandbox to compute, aggregate, transform, or analyze data. Assign module.exports to return a value. Snippets can call Dashback API routes using fetch() (e.g. await fetch("/article/2")). A sandboxed virtual filesystem (fs / vfs) is available to store and retrieve artifacts across runs.',
'4. find_files: List what is stored in the virtual filesystem (VFS). Call find_files({ skills: true }) to list SKILL files only.',
'5. read_file: Read a VFS file as text, windowed with seek/limit for large files.',
'6. patch: Create, overwrite, or search-and-replace text in a VFS file. This is the only text write path — it also creates SKILL files.',
'7. search_images: Search Google Images and store the candidates in the VFS as AVIF files.',
'8. analyze_image: Have a vision model look at up to 9 images (VFS paths or "image/<fullHash>") and return its judgement. This is how you "see" — never claim to have inspected an image yourself.',
'9. web_search: Search the web using Google Search. Returns organic results, snippets, and knowledge graph. Supports filtering by timePeriod ("day", "week", "month", "year") — always use "day" or "week" for news and recent events.',
'10. url_get: Fetch and scrape the content and markdown of a web page/URL.',
'11. pdf_read: Read and convert PDF or text files. Converts PDF to layout-preserved text ("txt") or renders pages into PNG images ("png") saved in the VFS images/ folder.',
'12. reset_vfs: Delete files, images, and artifacts stored in the virtual filesystem (VFS) to free up storage or clean up. SKILL files are exempt and survive.',
'13. rm: Delete a single file from the VFS. This is how a SKILL file is deleted on purpose, since reset_vfs always spares skills.',
'14. apisuggestionsorissues: Report missing API endpoints, bugs, or issues to be triaged. Automatically writes an entry to the company history table for review.',
'1. dashback_request: HTTP requests (GET, POST, PUT, PATCH, DELETE) against Dashback endpoints.',
'2. ask_user: Prompt user for clarification or decisions. Pauses execution until response.',
'3. javascript_execute: Sandboxed Node environment for computation, aggregation, and binary I/O. Assign module.exports. API: fetch("/path"). VFS: fs/vfs.',
'4. find_files: List VFS files. Set skills=true for SKILL files.',
'5. read_file: Read VFS text file (seek/limit windowed).',
'6. patch: Create, overwrite, or replace text in VFS files (only text write path; creates SKILL files).',
'7. search_images: Search Google Images; downloads AVIF candidates to VFS.',
'8. analyze_image: Inspect up to 9 images via vision model (VFS paths or "image/<hash>"). Always evaluate images through this tool.',
'9. web_search: Web search via Google. Filter with timePeriod ("day"/"week" for current events).',
'10. url_get: Fetch and scrape web page content/markdown.',
'11. pdf_read: Convert PDF/text to layout text ("txt") or VFS page images ("png").',
'12. reset_vfs: Delete VFS files/artifacts. Preserves SKILL files.',
'13. rm: Delete single VFS file (used to delete SKILL files or specific artifacts).',
'14. apisuggestionsorissues: Report missing API endpoints, bugs, or issues for triage.',
'',
'VIRTUAL FILESYSTEM:',
'- You have a persistent per-company VFS. Paths are relative to the VFS root and always use forward slashes, e.g. "labels/sheet1.png".',
'- To READ text: find_files (list) then read_file (content). To WRITE text: patch. Binary artifacts (images, PDFs, buffers) go through javascript_execute.',
'- Never invent a VFS path. Always confirm with find_files before reading or referencing a file.',
'- Persistent per-company VFS. Paths are root-relative forward-slash (e.g. "labels/sheet1.png").',
'- Text: list with find_files, read with read_file, write with patch. Binary/buffers: javascript_execute.',
'- Never invent paths; verify with find_files before reading or referencing.',
'',
'SKILL FILES:',
'- A SKILL file is a reusable instruction document: a file named "<slug>.SKILL" at the VFS root whose FIRST LINE is its title and whose remaining lines are markdown.',
'- SKILL file names may NOT contain "/" — skills are flat, at the root only. Use find_files({ skills: true }) to discover them; this listing is not affected by pagination or by how many artifacts exist.',
'- Read a skill with read_file BEFORE starting a task it covers, and follow it. If you learn a better procedure during the task, offer to write it back as a new SKILL file.',
'- SKILL files are exempt from reset_vfs: artifacts are disposable, skills are not. Never call reset_vfs expecting to clear skills, and never put throwaway output in a .SKILL file.',
'- To delete a skill (or any single file) use rm — it is the only way to remove one. Deletion is final.',
'- Reusable instruction docs: "<slug>.SKILL" at VFS root. First line is title, followed by markdown.',
'- Root-only (no "/"). Discover via find_files({ skills: true }).',
'- Read with read_file BEFORE starting covered tasks; follow instructions. Propose saving improved procedures as new SKILL files.',
'- Exempt from reset_vfs. Delete permanently via rm.',
'',
buildSkillIndex(vfsRoot),
'',
'IMAGES:',
'- You cannot see images yourself. To evaluate one, call analyze_image and report what the vision model says.',
'- Typical flow: search_images("query") → analyze_image({ paths: [...candidates], prompt: "…" }) → upload the winner with dashback_request.',
'- The `image/` folder holds database images and is READ-ONLY. Generated images go anywhere else (e.g. "images/…").',
'- Evaluate images via analyze_image; report model findings.',
'- Flow: search_images("query") → analyze_image({ paths: [...], prompt: "…" }) → upload via dashback_request.',
'- `image/` is READ-ONLY database storage. Write generated images elsewhere (e.g. "images/…").',
'',
'OPERATING GUIDELINES:',
'- Refer to the DASHBACK API ROUTES documentation below to find appropriate endpoints, parameters, and HTTP methods.',
'- SELLING PRICES: In article responses (GET /article/{sku} and returned article), the selling price field is named net (NOT price). The structure of entries in article.prices is { minmultiplier, net, taxRate, customergroupId }.',
'- GROSS vs NET PRICES: Users almost always mean GROSS (brutto, incl. VAT) prices when they specify a price (e.g. "1,60 €"), but the API stores and accepts NET prices only. Convert: net = gross / (1 + taxRate/100). Send net with high precision (at least 4–6 decimal places) so rounding back gives the exact gross. Calculate this in javascript_execute.',
'- DATA TYPES & WIRE FORMAT: Dashback strictly validates types (numeric values like weight, price, quantity, IDs must be numbers; flags must be booleans).',
'- SHOWING GENERATED IMAGES: Markdown images ARE supported and render inline in the chat. After writing an image into the virtual filesystem (e.g. `fs.writeFileSync("artifacts/example.png", buf)`), reference the exact path you wrote with standard markdown `![title](artifacts/example.png)`, which renders inline. Make the alt text descriptive. Other file types (PDF, JSON, CSV) are NOT rendered inline — link to them instead and offer them as a download.',
'- LABEL TEMPLATES (/labeltemplate): Label templates are NOT ZPL. content is Dashback\'s JSON label schema: { "objects": [ ... ] }. Do NOT write or suggest ZPL.',
'- Format responses clearly using markdown tables, bullet lists, or bold key-value pairs where appropriate.',
'- Endpoint reference: see DASHBACK API ROUTES below.',
'- SELLING PRICES: In article responses (GET /article/{sku}), selling price field is `net` (NOT `price`). Schema in `article.prices`: { minmultiplier, net, taxRate, customergroupId }.',
'- GROSS vs NET PRICES: User prices are almost always GROSS (incl. VAT); API requires NET. Formula: `net = gross / (1 + taxRate / 100)`. Compute in javascript_execute with high precision (4–6 decimals) to prevent rounding error.',
'- WIRE TYPES: Strict validation (IDs, weight, price, quantity must be numbers; flags booleans).',
'- DISPLAYING GENERATED IMAGES: Inline markdown images supported. Embed written VFS image paths: `![descriptive alt](artifacts/example.png)`. Other files (PDF, JSON, CSV) do not render inline; provide download links.',
'- LABEL TEMPLATES (/labeltemplate): Must use Dashback JSON schema `{ "objects": [ ... ] }`. Do NOT use ZPL.',
'- Format output with markdown tables, bullet points, or bold key-value pairs.',
...(routesDoc ? ['', 'DASHBACK API ROUTES:', routesDoc] : []),
].join('\n');