Skip to content

Raga Engine

The RAGA Engine handles concurrent multi-user inference with queue management, dynamic system prompt construction, and two primary output paths. This page complements the Architecture Overview with inference component details.

Engine Workflow

Engine Components

ComponentFunction
MAX_VLLM_RUNNING (QUEUE 4)User request queue, capping a maximum of 4 active requests concurrently
VLLM /metrics concurrent 2Runs model inference with a maximum of 2 parallel concurrency, with metrics monitoring
System Prompt BuilderDynamically constructs the system prompt based on the model's context window (32k / 128k / xx)
API Context (100,000 char)Injects external API search results into the context
Basic (10 items)Baseline workspace knowledge that is always included
OCR (3 items)Indexed documents via INDEX_OCR_KNOWLEDGE, INDEX_OCR_SUMMARIZE_CHUNK, INDEX_OCR_SUMMARIZE
RDBMS (10,000 char)Relational data, processed via Text-to-SQL
Audio (20 rows)Indexed audio transcripts via INDEX_AUDIO_KNOWLEDGE, INDEX_AUDIO_SUMMARIZE_CHUNK, INDEX_AUDIO_SUMMARIZE
History PromptPrevious conversation history within the session
Internal PromptRAGA's internal system instruction (not exposed to the user)
Time NowReal-time timestamp injection for temporal context
files_usersFiles uploaded by the user within the active session
PersonalizationUser-specific preferences and context
Image ContextMultimodal image content sent by the user
Text To APIConverts a user question into a structured API call
Text To SQLConverts a user question into a structured SQL query
LLM ProviderThe inference backbone is multi-provider, selected per request via the llm parameter: local (Qwen2.5-14B-Instruct-AWQ, default) and local_v2 (Qwen2.5-32B-Instruct-AWQ), both self-hosted via vLLM, plus OpenAI-compatible external providers — sambanova (Llama-3.3-70B), groq (Llama-3.3-70B-versatile), openai (gpt-4o-mini), and alibaba (Qwen-plus)
Multimodal ProviderA separate path for image analysis (Image Context): local (Qwen2.5-VL-32B-Instruct-AWQ) or alibaba (Qwen-VL-Max)

Summary

The RAGA Engine bounds concurrency through a queue (up to 4 active requests, 2 parallel inferences in VLLM), then dynamically assembles the system prompt from eleven context types — from baseline workspace knowledge to Text-to-SQL and Text-to-API results — before passing it to whichever LLM provider is configured (local Qwen-based vLLM by default, or an external provider such as SambaNova, Groq, OpenAI, or Alibaba as an alternative).