bytelyst-devops-tools/docs/llm-utility-workflows.md
Hermes VM 0e1905aa33
Some checks failed
pre-commit / pre-commit (push) Failing after 33s
docs: document local LLM utility workflows
2026-05-28 00:21:06 +00:00

157 lines
4.4 KiB
Markdown

# LLM Utility Workflows: FreeLLMAPI + MarkItDown
This VM has two private utilities for cheaper, safer, and more token-efficient AI workflows.
## FreeLLMAPI private fallback gateway
FreeLLMAPI is installed as a private OpenAI-compatible gateway for low-stakes fallback or optional use.
### Runtime
- App path: `/opt/freellmapi/app`
- Persistent database: `/var/lib/freellmapi/data/freeapi.db`
- Service: `freellmapi.service`
- Base URL: `http://127.0.0.1:3001/v1`
- Client env file: `/etc/freellmapi/client.env`
- Status helper: `freellmapi-status`
- Provider-key helper: `freellmapi-add-key`
The service is intentionally loopback-only. Do not expose it through Caddy or publish it on a Docker public interface; its dashboard/admin APIs are local-trust-only.
### Operations
```bash
systemctl status freellmapi.service --no-pager
freellmapi-status
journalctl -u freellmapi.service -n 100 --no-pager
systemctl restart freellmapi.service
```
Expected bind check:
```text
127.0.0.1:3001
```
There should be no `0.0.0.0:3001` or `:::3001` listener.
### Add provider keys
FreeLLMAPI is installed and healthy, but it needs provider keys before chat fallback is usable. Add keys without echoing secrets:
```bash
printf '%s' "$OPENROUTER_API_KEY" | freellmapi-add-key openrouter "main openrouter"
printf '%s' "$GEMINI_API_KEY" | freellmapi-add-key google "main gemini"
printf '%s' "$GROQ_API_KEY" | freellmapi-add-key groq "main groq"
```
For Cloudflare Workers AI, the expected FreeLLMAPI key format is:
```text
account_id:api_token
```
Check key counts without revealing secrets:
```bash
freellmapi-status
```
### Use from OpenAI-compatible clients
Load local client config:
```bash
set -a
source /etc/freellmapi/client.env
set +a
```
Then configure clients with:
- Base URL: `$FREELLMAPI_BASE_URL`
- API key: `$FREELLMAPI_API_KEY`
- Model: `auto`
Use this only for non-sensitive, low-stakes, or fallback workloads. Prompts still go to third-party providers selected by the gateway.
### Safety boundaries
- Do not send secrets, credentials, customer data, or sensitive incident details.
- Do not make this public.
- Treat it as opportunistic capacity, not an SLA-backed production dependency.
- Prefer primary paid/high-quality providers for serious Hermes coding/devops work.
## MarkItDown document-to-Markdown workflow
Microsoft MarkItDown is installed for token-efficient local document extraction before LLM analysis.
### Runtime
- Venv: `/opt/markitdown/venv`
- Direct CLI: `bytelyst-markitdown`
- Safe wrapper: `bytelyst-doc2md`
- Status helper: `markitdown-status`
- Hermes skill: `markitdown-document-workflow`
Installed support covers common local document formats: PDF, DOCX, PPTX, XLSX/XLS, HTML, CSV/JSON/XML/text-like inputs supported by MarkItDown.
### Standard conversion
```bash
bytelyst-doc2md /path/to/document.pdf -o /tmp/document.md
python3 -m json.tool /tmp/document.md.stats.json
```
The stats file records source size, SHA256, output size, token estimate, truncation status, and conversion time.
For large files:
```bash
bytelyst-doc2md /path/to/document.pdf -o /tmp/document.head.md --max-chars 50000
```
### Security behavior
`bytelyst-doc2md` is the preferred wrapper because it:
- refuses URLs by default,
- disables plugins by default,
- strips long base64 data URIs by default,
- records source SHA256 for local files,
- writes token estimates before content is loaded into an LLM context.
Use `--allow-url` only for trusted URLs after considering SSRF/local-file exposure. Prefer downloading the file with a vetted tool first, then converting the local path.
### Verification
```bash
markitdown-status
```
Expected output includes:
```text
bytelyst-doc2md: ok
source_sha256_present: True
markdown_heading_present: True
url_refusal: ok
```
## Hermes usage pattern
When a user sends a document or asks to analyze a Drive/PDF/Office file:
1. Load the `markitdown-document-workflow` skill.
2. Download or locate the file locally.
3. Convert with `bytelyst-doc2md`.
4. Read the stats JSON first.
5. Search/page the generated Markdown instead of loading the whole file.
6. Fall back to OCR tooling only when MarkItDown output is incomplete or image-only.
When a user asks to save LLM/provider cost:
1. Use primary Hermes provider for sensitive or high-stakes work.
2. Use FreeLLMAPI only for low-stakes fallback after provider keys have been added.
3. Confirm `freellmapi-status` shows at least one provider key before routing work to it.