admin/accounted
3
0
Fork 0
Code Issues 1 Pull Requests Actions Packages Projects Releases Wiki Activity

docs(self-host): document which AI flows run on any model vs need Anthropic-family (#1801)

Browse Source 
The assistant Q&A (both /chat and the docked sheet), document extraction, and
one-tap transaction categorization already run on the provider-agnostic
single-call path, so they work on an OpenAI-compatible or local model. A set of
specialized conversational flows still runs on the streaming run-turn runtime
and needs an Anthropic-family backend (Bedrock or the direct API); they 503 on
other backends. Spell out the boundary under AI Features and point at #1800 for
the remaining migration.

Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com>
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
Jakob Wennberg
2026-08-22 12:36:22 +02:00
committed by GitHub
co-authored by Jakob Wennberg Claude Opus 4.8
parent dc5079a912
commit 561e64afc7
1 changed files with 16 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/SELF-HOSTING.md
+16 -1
View File
@@ -219,7 +219,7 @@ AWS_REGION=eu-north-1 # default
Set both static AWS keys explicitly. The AI assistant's client can fall back to the standard AWS credential provider chain (instance profile, IRSA) when they are absent, but document extraction requires `AWS_ACCESS_KEY_ID` and `AWS_SECRET_ACCESS_KEY` and silently returns empty results without them.
**Option 3: any OpenAI-compatible endpoint.** Any server that implements the chat-completions API: a Swedish inference provider for a fully sovereign deployment, or a **local model** on the same machine (llama.cpp's `server`, Ollama's `/v1`, LM Studio, vLLM). Document extraction (receipts, invoices, HTML mail invoices) and the single-call assistant (`/api/agent/ask`) run here on any provider; the older streaming chat panel is being replaced by that single-call surface.
**Option 3: any OpenAI-compatible endpoint.** Any server that implements the chat-completions API: a Swedish inference provider for a fully sovereign deployment, or a **local model** on the same machine (llama.cpp's `server`, Ollama's `/v1`, LM Studio, vLLM). Document extraction (receipts, invoices, HTML mail invoices), the assistant's question-and-answer (on both `/chat` and the docked assistant sheet, via `/api/agent/ask`), and one-tap transaction categorization run here on any provider. A set of specialized conversational flows still needs an Anthropic-family backend; see **What runs on any model** below.
```bash
AI_BASE_URL=http://localhost:11434/v1 # the endpoint's OpenAI-compatible base URL (here: a local Ollama)
@@ -235,6 +235,21 @@ Three things about such endpoints are declared rather than probed, because the a
- `AI_PDF_MODE` defaults to `rasterize` here: most such endpoints have no PDF input, so the first `AI_PDF_MAX_PAGES` pages (default 4) are rendered to images with poppler's `pdftoppm`, which the self-host image installs (`apk add poppler-utils`, the only system package beyond the base image; page images are written to `/tmp`, a tmpfs in `docker-compose.yml`). If the binary is missing, PDFs are skipped with `pdf_rasterizer_missing` rather than failing; `AI_PDF_RASTERIZER_BIN` points at a non-standard install. A provider that accepts the OpenAI `file` content part can use `AI_PDF_MODE=native`.
- `AI_STRICT_JSON=true` asks for `response_format: json_schema` on providers that enforce it. The default (JSON answered in prose, then parsed and validated) works on every model and is what hosted runs.
#### What runs on any model
Most AI surfaces run on any of the three backends above, an OpenAI-compatible or local model included:
- **Document extraction**: receipts, invoices, and HTML mail invoices.
- **The AI assistant's question-and-answer**, on both `/chat` and the docked assistant sheet (`/api/agent/ask`), including its read-only ledger tools.
- **One-tap transaction categorization** (`/api/agent/categorize`): the deterministic-candidates then model-select cascade shown on the transactions page.
A few **specialized conversational flows** still run on the older streaming runtime (`/api/agent/invoke`), which requires an Anthropic-family model (AWS Bedrock or the direct Anthropic API). On an OpenAI-compatible or local backend these specific actions return `503`; everything above keeps working. They are:
- **Operation-staging flows** (they draft a change for you to approve): the invoice-inbox "Fraga assistenten" categorize, bulk-book, invoice draft, supplier-invoice review, and verifikat draft.
- **Context-bound assistant helpers**: VAT review, KPI explanation, settings help, the bokslut step-through, and onboarding.
Migrating these to the single-call surface, so a fully local deployment covers them too, is tracked in [#1800](https://github.com/erp-mafia/accounted/issues/1800).
Optional model overrides, in any setup:
```bash