Z
Ceteris-Paribus Pro® Console
Ceteris-Paribus Pro® · Tumart Vector v3.5
API Configuration
Ceteris-Paribus Pro® · runtime secrets

Values entered here are saved to disk and take effect immediately — no restart needed. They override Replit Secrets for this session and all future sessions.

Ceteris-Paribus Pro® — Generation
AI_PROVIDER_BASE_URL
AI_GENERATION_MODEL
AI_PROVIDER_API_KEY
Tumart Vector v3.5 — Embedding Provider
AI_EMBEDDING_MODEL — default: text-embedding-3-small
EMBEDDING_BASE_URL — leave blank to use OpenAI
EMBEDDING_API_KEY — leave blank to fall back to OPENAI_API_KEY
EMBEDDING_DIM — default: 1536
⚠ Changing the dimension only takes effect for new corpus tables. If you change this value you must clear the entire vector corpus and re-ingest all documents.
OpenAI fallback key
OPENAI_API_KEY — not required when EMBEDDING_API_KEY is set
Batch Embeddings

When a single ingest job exceeds the threshold, the Sovereign Ingestion Gateway switches to the Fireworks Batch API — lower cost, no rate-limit pressure. Requires EMBEDDING_BASE_URL. Falls back silently to sync on any error.

BATCH_EMBED_ENABLED
BATCH_EMBED_THRESHOLD — default: 100 chunks
Crawler / Extractor — Tender Extraction LLM

Used by the Sovereign Semantic Distillation layer to extract structured tender notices from page changes and PDFs. Falls back to the generation provider when blank.

CRAWLER_BASE_URL
CRAWLER_MODEL — default: accounts/fireworks/models/llama-v3p1-70b-instruct
CRAWLER_API_KEY — leave blank to fall back to AI_PROVIDER_API_KEY
Alerts — Relevance Gating

Tender notices whose composite relevance score meets or exceeds this threshold automatically create an entry in the Alerts panel. Score = 60% extraction confidence + 40% keyword weight.

0 (all)50100 (highest only)
ALERT_THRESHOLD — default: 60
Retrieval — Reranker

When enabled, Tumart Vector v3.5 retrieves 20 cosine candidates and a cross-encoder reranker re-scores them against the exact query before passing the top 4 to the generation model. Uses the same provider and API key as generation. Falls back to cosine order transparently when the provider does not support /v1/rerank.

RERANKER_ENABLED — default: true
RERANKER_MODEL — default: nomic-ai/nomic-rerank-v1-5
Vision / PDF Ingestion

PDF files are rendered page-by-page at 150 dpi and sent to a vision language model for text extraction. By default the vision pipeline reuses AI_PROVIDER_BASE_URL and AI_PROVIDER_API_KEY. Set the fields below to route PDF extraction to a different endpoint (e.g. a cheaper 11B vision model on Fireworks).

VISION_MODEL — falls back to llama-v3p2-11b-vision-instruct
VISION_BASE_URL — falls back to AI_PROVIDER_BASE_URL
VISION_API_KEY — falls back to AI_PROVIDER_API_KEY
Active runtime configuration
Ceteris-Paribus Pro®
Provider base URL
Tumart Vector v3.5 endpoint
PDF ingestion
Add from URL
Paste one or more public web pages (one URL per line) and we'll fetch, extract, and ingest them.
Sovereign Ingestion Gateway
Drop .txt, .md, .json, .pdf, .docx, or image files into uploaded_docs/, or upload directly below.
No files chosen
Filename / source Vectors Last ingested
loading…
Total vectors in Tumart v3.5: 0
Research Agent
Autonomous web crawler · LLM-scored proposals
Advanced crawl parameters
101 = disabled
Surveillance Nodes
Tier Node Actions
Open panel to load…
Tender Notices
Reference Title / Entity Score
Open panel to load…
Alerts
Open panel to load…
Answer

Ask a question about your ingested documents to see a streamed, grounded answer.

Retrieved context

Source snippets will appear here as soon as a query is dispatched.