Project Synapse

Advanced Multi-Pipeline AI Research & Analysis Platform

Part of the Xere AI unified intelligence system, developed by EdgeXene LLC

Mission Statement

Project Synapse, developed under eklypse as part of the Xere AI unified intelligence system, delivers AI assistance built for something more useful than writing bad haikus; it's designed for real personal and research work. It's built to be more than just another chatbot: think personal research buddy meets brainy sidekick. With transparent multi-stage reasoning (no black-box mumbo jumbo), real-time data plugged in, and security solid enough to make a lawyer sleep at night, Synapse's aim is to deliver reliable, citation-backed insights.

Whether it's helping with independent legal digging, breaking down business strategy, or just stress-testing wild ideas, the mission is simple: keep improving the platform while poking at the frontier of agentic RAG, so eventually, it won't just help with research, it'll run autonomous research workflows on its own (without asking for coffee breaks) -- like a tireless digital colleague.

Built for Real Work • Transparent by Design • Pushing Toward Agentic RAG

Platform Overview

Project Synapse is a multi-pipeline AI research and analysis platform with 9 specialized modes and a full document management system. Built on Node.js with Together AI serverless inference, Qdrant vector search, and 17+ integrated data sources.

9
Specialized Modes
192GB
DDR5 ECC RAM
17+
Data Integrations
6
Export Formats

Specialty Mode Pipelines

All specialty modes use GLM-5.1 for Stage 1 intelligent dispatching with semantic API routing. Each mode is a multi-stage pipeline powered by Together AI serverless inference with automatic fallback cascading. Each mode is detailed in its own section below.

Consultant

Business strategy and analysis: a 3-stage pipeline that dispatches to economic and market data sources, then performs deep analysis and strategic synthesis.

Max Tokens: 78,000

Stages: 3

Data Sources: FRED, World Bank, Polygon.io, SEC EDGAR, RAG, Tavily, NewsAPI

Stage 1: zai-org/GLM-5.1 (6k, temp 0.5) -- Dispatch + API routing
↓
Stage 2: openai/gpt-oss-120b (48k, temp 1.0) -- Deep analysis
↓
Stage 3: deepcogito/cogito-v2-1-671b (24k, temp 0.8) -- Strategic synthesis

Automatic failover (if a model is temporarily unavailable): cogito-v2-1-671b → Qwen3-235B → gpt-oss-120b

Legal

Legal research and compliance: a 3-stage pipeline that routes to legal data sources, reasons over the law, and produces a cited final synthesis.

Max Tokens: 81,000

Stages: 3

Data Sources: CourtListener, UK Legislation, Congress.gov, SEC EDGAR, RAG, Tavily, NewsAPI, World Bank

Stage 1: zai-org/GLM-5.1 (6k, temp 0.4) -- Dispatch + API routing
↓
Stage 2: deepcogito/cogito-v2-1-671b (48k, temp 0.7) -- Legal reasoning
↓
Stage 3: Qwen/Qwen3-235B-A22B-Instruct-2507-tput (27k, temp 0.6) -- Final synthesis

Automatic failover (if a model is temporarily unavailable): Qwen3-235B → gpt-oss-120b (Stage 2 cogito-v2-1-671b → Qwen3-235B → gpt-oss-120b)

Technical

Engineering and code generation: a 3-stage pipeline with an optional auto-triggered Stage 4 for extended code implementation.

Max Tokens: 93,000 base (+ 51k code gen)

Stages: 3 (+ optional Stage 4)

Data Sources: RAG, Tavily, arXiv, Wikipedia, NASA, NewsAPI

Stage 1: zai-org/GLM-5.1 (6k, temp 0.5) -- Intelligent dispatching
↓
Stage 2: Qwen/Qwen3.5-397B-A17B (36k, temp 0.6) -- Deep technical analysis
↓
Stage 3: MiniMaxAI/MiniMax-M3 (51k, temp 0.4) -- Code generation & synthesis
↓
Stage 4 (auto-trigger): MiniMaxAI/MiniMax-M3 (51k) -- Extended code generation

* Stage 4 auto-triggers when code implementation is requested

Automatic failover (if a model is temporarily unavailable): Code-gen (MiniMax-M3) → Kimi-K2.7-Code → gpt-oss-120b; Stage 2 ↔ MiniMax-M3; Stage 3 ↔ Qwen3.5-397B

Creative Writing

Long-form and short-form creative writing: a 2-stage pipeline with a choice of two Stage 2 styles.

Max Tokens: 83,000

Stages: 2

Styles: 2 writing modes

Data Sources: RAG, Tavily, Wikipedia, NASA, NewsAPI

Stage 1: zai-org/GLM-5.1 (6k, temp 0.6) -- Creative direction
↓
Stage 2a: Qwen/Qwen3.5-397B-A17B (77k, temp 0.9) -- Polished Narrative
or
Stage 2b: meta-llama/Llama-3.3-70B-Instruct-Turbo (77k, temp 0.9) -- Dynamic Shorts

* Polished Narrative for literary fiction, memoirs, structured long-form. Dynamic Shorts for flash fiction, dialogue-heavy scenes, experimental drafts.

Automatic failover (if a model is temporarily unavailable): Polished Narrative (Qwen3.5-397B) → Qwen3-235B → gpt-oss-120b; Dynamic Shorts (Llama-3.3-70B) → Qwen3.5-397B → gpt-oss-120b

Image Generator

Dedicated image generation with a choice of two models and a set of art style presets.

Type: Dedicated image generation mode

Models: 2 options

Data Sources: None (image synthesis only)

Imagen 4.0 -- Fast generation for quick iterations
or
FLUX.1.1-pro -- High-quality image synthesis

* 12 art style presets per model (Anime, Photorealistic, Digital Art, Futuristic, Landscape, Oil Painting, Concept Art, Watercolor, 3D Render, Pixel Art, Minimalist, Dark Fantasy)

Ask Anything

Real-time web search with Perplexity-style streaming results. Supports speed modes (fast, balanced, deep), automatic model escalation based on context size, and location-aware local queries.

Default Model: zai-org/GLM-5.1 (temp 0.3)

Large Context (>25k tokens): Escalates to Qwen/Qwen3-235B-A22B-Instruct-2507-tput

Math/Logic Queries: Escalates to Qwen/Qwen3-235B-A22B-Instruct-2507-tput

Web Search: Tavily (Brave fallback) with perspectives

Citation Styles: OSCOLA (default), MLA

Data Sources: Tavily, Wikipedia, RAG

Tavily Web Search, Brave fallback (fast/balanced/deep)
↓
Batch Content Scraping (source extraction)
↓
LLM Synthesis with citations (model selected by query complexity)

Automatic failover (if a model is temporarily unavailable): selected model → gpt-oss-120b

Post-escalation: if a gpt-oss-120b answer scores low confidence (< 55%), it is regenerated once with deepcogito/cogito-v2-1-671b for a higher-confidence response

Agentic Research

Long-form research paper generation with 4 stages (+ optional agentic deep research). Supports configurable word counts from 3,000 to 30,000 words with real-time progress tracking via SSE.

Word Counts: 3k, 5k, 10k, 15k, 20k, 25k, 30k

Citation Styles: OSCOLA (default), APA, MLA

Export Formats: PDF, DOCX, Excel (XLSX), LaTeX, RTF, Markdown

Data Sources: RAG (always), Tavily, arXiv, CourtListener, UK Legislation, Congress, FRED, World Bank, Polygon.io, SEC EDGAR, Alpha Vantage, NASA, Wikipedia, NewsAPI

Stage 1: Planning -- Outline generation and citation estimation
↓
Stage 2: Research -- Source gathering and deep research
↓
Stage 2b (optional): Agentic Deep Research -- Iterative gap filling (up to 3 iterations, 85% coverage threshold)
↓
Stage 3: Synthesis -- Comprehensive paper body with citations
↓
Stage 4: Quality Assurance -- Citation verification, consistency checking, polish

Options: Enable/disable RAG, web search, legal APIs, agentic research, and LaTeX generation per paper.

Automatic failover (if a model is temporarily unavailable): each stage's model follows the standard cascade — GLM-5.1 → gpt-oss-120b; cogito-v2-1-671b / Qwen3.5-397B → Qwen3-235B → gpt-oss-120b

Monitor

Continuous web monitoring powered by Brave Search with Together AI summaries. Define a natural-language query once, choose a cadence, and Synapse polls the web for new events matching your query. Events persist in your chat history, even while you are signed out.

Max Tokens: 400 (per-event summary)

Stages: 2 (search + summarize)

Data Sources: Brave Search

Cadences: Every 6, 8, or 12 hours, or every 1, 3, 5, or 7 days

Per-user limit: 3 active monitors

Duration: Each monitor runs for up to 21 days, then stops automatically

History retention: Events remain visible for 365 days after expiry

Stage 1: Brave Search -- server-side poll for new results matching the query
↓
Stage 2: meta-llama/Llama-3.3-70B-Instruct-Turbo (400 tokens, temp 0.3) -- summarize new results into an event

How it works: A background poller on Synapse's servers runs each monitor on its chosen cadence, searching the web via Brave Search and summarizing new results with Llama-3.3-70B. Detected events are persisted to your account so you can review them next time you sign in -- no action needed on your end.

Financial

A four-tab markets workspace (Stock, ETF, Economic, News) for equities, funds, economic indicators, and global markets, with an AI analysis layer. Market data comes from third-party providers; quotes and figures are delayed, not real-time, and are for research and informational use only (not trading). Prices and figures are never AI-generated.

Max Tokens: 8,000 (AI analysis)

AI: single-pass MiniMax-M3 (DeepSeek-V4-Pro fallback)

Tabs: Stock (quote + chart + AI analysis + news), ETF (common-ETF picker + quote + chart + AI analysis), Economic (FRED indicators), News (Global Markets + Index News)

Data Sources: Polygon.io & Alpha Vantage (quotes, history), FRED (economic indicators), web search (global indices & headlines)

Data layer: Polygon.io, Alpha Vantage & FRED -- delayed market data and economic indicators (no AI)
↓
AI analysis: MiniMaxAI/MiniMax-M3 (8k, temp 0.6) -- one-pass buy-side analysis with directional view; DeepSeek-V4-Pro fallback

How it works: Quotes and charts come directly from the market-data providers; the Global Markets card and broad headlines are sourced via web search (the data feed has no real global-index coverage). The AI layer reasons over the data to produce a written analysis with a directional view (no explicit buy/sell or price targets) — it never generates prices or figures, which always come from the delayed feed.

Grand Library & RAG System

Centralized document management system with layout-aware PDF processing, vector search, and agentic retrieval.

Document Management

Collections Personal, public, and legal document collections with custom categories and tags
Supported Formats PDF, DOCX, TXT, and spreadsheets (XLSX, XLSM, XLS, CSV) up to a 50MB upload limit
PDF Processing pdfjs-dist for layout detection + Qwen3.5-397B vision model (Kimi-K2.7-Code fallback) for tables, charts, and figures
Spreadsheet Intelligence Every sheet is parsed to typed data with cell formulas surfaced alongside computed values. Ask calculational questions about a pinned spreadsheet ("total revenue", "average units per region") and the answer is computed exactly with pandas in a local sandbox — your data never leaves the server. Best in Consultant or Legal mode.
ArXiv Auto-Fetch Research and Technical modes automatically download cited ArXiv papers into the Grand Library

Agentic RAG Pipeline (3-Stage)

Stage 1: Dispatcher

Model: zai-org/GLM-5.1
Role: Analyzes query intent and extracts search parameters. Routes to RAG Tool, Web Search, or Specialized APIs.
Routing: Semantic routing across 17 data sources via structured JSON tool selection

↓
Stage 2: Analyzer

Model: deepcogito/cogito-v2-1-671b
Role: Evaluates quality and sufficiency of retrieved results, calculates confidence scores, determines if supplementation or self-correction is needed

↓
Stage 3: Synthesizer

Model: openai/gpt-oss-120b
Role: Creates comprehensive final answer with OSCOLA-style [GL-n] citations and bidirectional navigation

Automatic failover (if a model is temporarily unavailable): GLM-5.1 → gpt-oss-120b; cogito-v2 → Qwen3-235B → gpt-oss-120b; gpt-oss-120b → Qwen3-235B. Self-correction loops re-run retrieval when Stage 2 confidence is low.

Infrastructure:

  • Embeddings: intfloat/multilingual-e5-large-instruct (1024 dimensions, multilingual)
  • Reranker: Optional second-stage reranking (configurable via RERANK_MODEL; disabled by default)
  • Vector DB: Qdrant (multiple collections: legal, financial, business, manufacturing, technical, public)
  • Knowledge Graph: Neo4j with concept-aware multi-hop retrieval (CITES, RELATES_TO, PART_OF, SUPPORTS, APPLIES_TO edges). Entity-based expansion augments vector search with graph traversal for citation-aware answers. Nightly backfill cron extracts new relationships from ingested documents.
  • Caching: Redis (1-hour TTL for embeddings)
  • Orchestration: LangGraph with conditional routing, graph enrichment node, and self-correction loops
  • Observability: LangSmith run tracing across the dispatcher, analyzer, and synthesizer stages plus tool execution and self-correction loops
  • Continuous Web Monitoring: Server-side poller (Brave Search + Together AI summaries, per-user monitors, 21-day max duration)

Token Allocation by Specialty Mode

GLM-5.1 Dispatching: All specialty modes use zai-org/GLM-5.1 (6k tokens) at Stage 1 for semantic API routing across 17 data sources. Ask Anything uses a separate model selection path with automatic escalation based on context size and query type. Agentic Research uses a dynamic, word-count-driven budget (1.3 tokens/word) — the figure shown is the maximum output budget for a 30,000-word paper; smaller papers allocate proportionally less. Image Generator is omitted from this chart because it produces images rather than text tokens.

Active Data Integrations

GLM-5.1 uses semantic understanding to automatically select the most relevant APIs for each query. The badges below show primary APIs per mode, but GLM-5.1 can route to any source based on query context.

Mode Primary Data Sources
Ask Anything Tavily Wikipedia RAG
Consultant FRED World Bank Polygon.io SEC EDGAR RAG Tavily NewsAPI
Legal CourtListener UK Legislation Congress.gov SEC EDGAR RAG Tavily NewsAPI World Bank
Technical RAG Tavily arXiv Wikipedia NASA NewsAPI
Agentic Research RAG (always) Tavily arXiv CourtListener UK Legislation Congress FRED World Bank Polygon.io SEC EDGAR Alpha Vantage NASA Wikipedia NewsAPI
Creative RAG Tavily Wikipedia NASA NewsAPI
Image Generator None (image synthesis only)
Monitor Brave Search
Financial Polygon.io Alpha Vantage FRED Tavily (global indices)

Intelligent Features

  • GLM-5.1 Semantic Dispatching: Structured JSON tool selection for intelligent API routing across all specialty modes
  • Complexity Detection: Queries classified as LOW/MEDIUM/HIGH based on depth, breadth, reasoning, ambiguity, and expertise dimensions
  • OSCOLA Citation System: All sources return OSCOLA-formatted citations -- RAG ([GL-N]), Web ([WEB-N]), Government ([GOV-N]), ArXiv ([ARXIV-N]) -- with bidirectional navigation
  • Model Fallback Cascading: Each model has a defined fallback chain -- GLM-5.1 to gpt-oss-120b, cogito-v2-1-671b to Qwen3-235B then gpt-oss-120b, Qwen3.5-397B to Qwen3-235B then gpt-oss-120b, MiniMax-M3 to Kimi-K2.7-Code then gpt-oss-120b
  • Multi-Jurisdiction Legal APIs: US (CourtListener, Congress.gov, SEC EDGAR), EU (EUR-Lex), UK (legislation.gov.uk)
  • ArXiv Auto-Download: Automatic paper retrieval with OSCOLA academic citations, ingested into Grand Library for future RAG queries
  • Together AI Key Rotation: Multiple API keys with automatic rotation for load distribution
  • Conversation Retention: 365-day history across all modes with per-mode session context

Platform Capabilities

  • GLM-5.1 intelligent dispatching -- Semantic API routing across 17 data sources via structured JSON tool selection
  • Multi-stage reasoning pipelines -- Up to 4 stages per specialty mode with automatic fallback cascading
  • Real-time SSE streaming -- Chat and research modes (Consultant, Legal, Technical, Creative, Ask Anything, Agentic Research) stream responses live with progress tracking; Image Generator, Monitor, and Financial return results directly
  • Legal & regulatory research -- Multi-jurisdiction APIs with OSCOLA citation formatting
  • Academic research -- arXiv, NASA, Wikipedia with auto-download to Grand Library
  • Citation systems -- OSCOLA (legal), APA (business/research), MLA (academic)
  • Export formats -- PDF, DOCX, Excel (XLSX), LaTeX, RTF, Markdown for research papers
  • Image generation -- Imagen 4.0 (fast) and FLUX.1.1-pro (quality) with 12 art style presets each
  • Professional-grade security -- Input sanitization, rate limiting (120 req/min), IP-based deduplication, threat detection
  • Conversation memory -- 365-day retention with session-based context per mode

Project Synapse

A product of Xere AI by EdgeXene LLC

Privacy Notice & Disclaimer Login

15442 Ventura Blvd., Ste 201-2525, Sherman Oaks, CA 91403

contact@edgexene.io

Last updated: June 16, 2026