Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/TrinaxCode/TrinaxAI/llms.txt

Use this file to discover all available pages before exploring further.

TrinaxAI ships five hardware profiles that automatically tune every performance-sensitive setting to your machine’s capabilities. The installer detects your total RAM and selects the appropriate profile — no manual tuning needed for most users. The profile is the master switch. It sets the model fleet, context window size, embedding worker count, chunk sizes, retrieval depth, and keep-alive durations in a single step.

​
How Profiles Work

When you run ./install.sh, the installer reads your total system RAM and writes TRINAXAI_PROFILE=<profile> into your .env file. From that point on, every restart reads the profile and applies its defaults. You can override the auto-detected profile at any time: 
./install.sh --profile ultra
./install.sh --profile max
./install.sh --profile 8gb
Run trinaxai doctor to see the currently active profile, all resolved model assignments, context window size, embed workers, and whether each Ollama model is downloaded. This is the fastest way to verify your configuration is correct.

​
Profile Reference

The table below shows the exact defaults applied by each profile, drawn directly from config.py.
Setting4gb / 8gb16gb (default)max / 32gbultra / 64gb
RAM target~4–8 GB~16 GB~32 GB64 GB+ or GPU workstation
MODEL_GENERALllama3.2:1bllama3.2:3bllama3.2:3bllama3.2:3b
MODEL_CODEqwen2.5-coder:1.5bqwen2.5-coder:3bqwen2.5-coder:3bqwen2.5-coder:3b
MODEL_DEEPqwen2.5-coder:1.5bqwen2.5-coder:3bqwen2.5-coder:7bqwen2.5-coder:14b
MODEL_FASTllama3.2:1bllama3.2:3bllama3.2:3bllama3.2:3b
Embed presetlitebalancedbalancedbalanced
Embed modelnomic-embed-textbge-m3bge-m3bge-m3
Embed dims768102410241024
NUM_CTX20484096819216384
EMBED_WORKERS1246
EMBED_BATCH18816
KEEP_ALIVE (fast mode)0s10m30m60m
EMBED_KEEP_ALIVE10m15m30m30m
CHUNK_SIZE (balanced)1024102410241536
CHUNK_OVERLAP (balanced)150150150220
SIMILARITY_TOP_K (fast)3456
FUSION_CANDIDATES (fast)681220
MODEL_GENERAL, MODEL_CODE, and MODEL_FAST share the same defaults across 16gb, max, and ultra. The key differentiator between profiles is MODEL_DEEP — complex queries (refactoring, architecture, multi-file analysis) escalate to a progressively larger model.

​
Profile Deep Dives

This profile prioritises staying within RAM. Everything is sized for machines where the OS itself consumes 3–4 GB.
  • Models: llama3.2:1b for chat, qwen2.5-coder:1.5b for code and deep tasks — the smallest Ollama-compatible options.
  • Embeddings: nomic-embed-text (768 dims, lite preset) — faster to load and query than bge-m3.
  • Context: 2048 tokens — enough for focused queries without competing with the model’s weights.
  • Concurrency: 1 embed worker, batch size 1 — serialised embedding to avoid OOM during indexing.
  • Keep-alive: 0s for LLM — model is unloaded after each response to free RAM for embeddings and the OS.
TRINAXAI_PROFILE=8gb
The factory default. Balanced for everyday coding, documentation search, and RAG queries.
  • Models: llama3.2:3b for chat, qwen2.5-coder:3b for code and deep analysis (the max/32gb profile upgrades deep analysis to qwen2.5-coder:7b).
  • Embeddings: bge-m3 (1024 dims, balanced preset) — multilingual, high-quality, 8K token context.
  • Context: 4096 tokens — comfortably fits a system prompt + 4 source chunks + response.
  • Concurrency: 2 embed workers, batch size 8 — parallel embedding without crowding the LLM.
  • Keep-alive: 10m in fast mode — models stay warm for several minutes between queries.
TRINAXAI_PROFILE=16gb
Unlocks the 7b deep model and wider context. Suitable for large codebases and multi-file analysis.
  • Models: Same general/code fleet as 16gb, but MODEL_DEEP is upgraded to qwen2.5-coder:7b — the first profile to unlock the larger deep-analysis model.
  • Context: 8192 tokens — fits longer documents and more retrieved chunks.
  • Concurrency: 4 embed workers, batch size 8 — faster bulk indexing.
  • Keep-alive: 30m — models stay warm for extended work sessions.
  • Retrieval: TOP_K=5, FUSION_CANDIDATES=12 — broader retrieval sweep.
TRINAXAI_PROFILE=max
Maximum quality. Designed for ML workstations, server-grade machines, and high-end consumer GPUs.
  • Models: qwen2.5-coder:14b for deep tasks — the best local code model available at time of writing.
  • Context: 16384 tokens — fit entire files, multiple retrieved documents, and long conversation history.
  • Concurrency: 6 embed workers, batch size 16 — rapid bulk indexing of large repositories.
  • Keep-alive: 60m — all models stay hot indefinitely during a session.
  • Chunking: Larger chunks (CHUNK_SIZE=1536, CHUNK_OVERLAP=220) capture more context per embedding.
  • Retrieval: TOP_K=6 (up to 8 in quality mode), FUSION_CANDIDATES=20–32 — exhaustive retrieval.
TRINAXAI_PROFILE=ultra
Valid aliases for ultra: gpu, 64gb, 64g, 4090, rtx, workstation.

​
Performance Modes

Performance mode is a secondary dial that fine-tunes chunk sizes and retrieval depth within a profile. It is independent of the profile — any profile can run in any mode.
ModeChunk sizeOverlapTOP_KFUSION_CANDIDATESBest for
fast (default)896 tokens96 tokens3–6 (by profile)6–20 (by profile)Snappy responses, frequent queries
balanced1024 tokens150 tokens4–6 (by profile)8–20 (by profile)General daily use
quality1024 tokens150 tokens5–8 (by profile)12–32 (by profile)Deep research, broad retrieval
TRINAXAI_PERFORMANCE_MODE=fast      # default — fastest responses
TRINAXAI_PERFORMANCE_MODE=balanced  # middle ground
TRINAXAI_PERFORMANCE_MODE=quality   # best retrieval precision
fast mode also uses smaller chunk overlaps (96 vs 150) and shorter code overlaps (8 vs 12 lines). This reduces index size and speeds up retrieval at a minor cost to cross-chunk context continuity.

​
Embedding Presets by Profile

The embedding preset is selected automatically from the profile but can be overridden independently.
PresetModelDimsToken ctxProfile default
balancedbge-m31024819216gb, max, ultra
litenomic-embed-text76820484gb, 8gb
fastall-minilm384512(manual override only)
To use a lighter embedding model on a 16gb machine (e.g., to reserve RAM for larger LLMs): 
TRINAXAI_PROFILE=16gb
TRINAXAI_EMBED_PRESET=lite
Changing the embed preset after indexing will produce vectors of a different dimensionality. You must re-index your documents after changing TRINAXAI_EMBED_PRESET or TRINAXAI_EMBED_DIMS, or queries will return poor results.

​
Overriding Individual Settings

A profile is a set of defaults, not a lock. Every value it sets can be overridden by the corresponding environment variable. The profile applies first; explicit variables override it. 
# Use the 16gb profile but with a larger context window and the deep model always warm
TRINAXAI_PROFILE=16gb
TRINAXAI_NUM_CTX=8192
TRINAXAI_KEEP_ALIVE=30m
TRINAXAI_EMBED_WORKERS=4
Switching profile mid-project: 
# Update .env
echo "TRINAXAI_PROFILE=max" >> .env

# Restart services
trinaxai stop
trinaxai start

# Verify the new profile is active
trinaxai doctor

Build docs developers (and LLMs) love

Get started for freeTalk to us