Configuration

Configuration — settings.json reference, models and backends

All settings live in settings.json in the project root — gitignored, local to each install, shared between the CLI and the Web UI. Edit it directly, or use menu option 8 (advanced settings) / option 5 (models).

A corrupt or partially-written settings.json never breaks a session: every field is type-coerced on load and falls back to its default.

---

Full settings reference

Key	Default	Description
model_a	llama3:latest	Primary generation model
model_b	qwen2.5:7b	Second model for parallel runs
synthesis_model	qwen2.5:7b	Model that merges two outputs
temperature	0.3	Sampling temperature, clamped to 0.0–2.0
timeout	600	Seconds allowed per model call
ollama_url	local Ollama endpoint	Where to send requests
techniques	[1,5,8,10,12,14,18,25,40,47,108,121,125,147,153]	Active technique IDs
use_web	true	Web enrichment on/off
stream	true	Real-time token streaming
output_mode	"full"	"quick" = enhanced prompt · "full" = 12-section manifest
use_pre_processor	true	Run the pre-processing step
pre_processor_model	""	Model for pre-processing ("" = use model_a)
preprocessor_mode	"llm"	"llm" = full model rewrite · regex-only mode is faster and makes no extra call
max_web_pages	1	Search results to fetch when enrichment is on
summarize_web_pages	false	Summarize fetched pages before injecting them
draft_mode	false	Generate only sections 1–2 of a Full manifest
use_result_cache	true	Serve identical repeat requests from cache/
anonymize_pii	false	Redact personal information before it reaches any model
encrypt_memory	false	Encrypt session memory at rest (needs cryptography)
backend_type	"ollama"	"ollama" or "openai_compatible"

Every one of these has a CLI flag equivalent that overrides it for a single run — see CLI Reference.

---

Choosing models

Role	What matters
model_a	Your workhorse. Quality/speed balance for the bulk of generations
model_b	Should differ in temperament from A — a creative model next to a systematic one gives synthesis real material to work with
synthesis_model	Favor instruction-following and long-context handling over raw creativity; it's merging, not inventing
pre_processor_model	A small fast model is ideal — this step only restructures your input

A practical setup: an 8B general model as A, a different-family 7B as B, and the stronger of the two as synthesis, with a 3B model for pre-processing.

---

Result caching

Identical requests are served from cache/ instantly, skipping the model call entirely. This makes iterating on flags cheap — change --mode or a technique set and only the genuinely new call costs time.

• Bypass for one run: --no-cache
• Disable permanently: "use_result_cache": false
• Clear it: delete the cache/ directory

---

Alternate backends

Ollama is the default, but any OpenAI-compatible local server works — LM Studio, GPT4All's server mode, text-generation-webui's OpenAI extension:

python3 prompt_expert_enhance.py generate "TASK" \
  --backend openai_compatible \
  --ollama-url http://localhost:1234/v1

Or persist it:

{
  "backend_type": "openai_compatible",
  "ollama_url": "http://localhost:1234/v1"
}

Both the chat-completions and legacy completions response shapes are handled, streaming included. Full notes: examples/alternate_backend.md.

The backend must run on the same machine. Only loopback URLs are accepted — localhost, 127.0.0.1 or ::1. Remote endpoints are rejected by design.

---

Session memory

Past runs are recorded in memory/ and injected as context on later runs, so a sequence of related tasks stays coherent instead of restarting cold each time.

Action	How
View	python3 prompt_expert_enhance.py memory view (or menu option 9)
Clear	python3 prompt_expert_enhance.py memory clear (or menu option 10)
Skip for one run	--no-memory, or the /neuf metacommand
Encrypt at rest	"encrypt_memory": true — requires pip install cryptography

With encrypt_memory on but cryptography missing, memory stays plaintext and the tool warns once instead of failing.

---

Where files live

Directory	Contents	Tracked by git?
outputs/	Generated manifests and prompts	No
memory/	Session history	No
cache/	Cached results	No
settings.json	Your configuration	No

For the compiled standalone app, these move to the standard per-OS user data directory instead — app bundles are read-only (and code-signed on macOS), so writable data can't live beside the binary. See Installation.

---

Next: CLI Reference — the per-run flag equivalents · Architecture — what each setting actually changes · Troubleshooting.