Muffakir Composer // New Search Run

source Retrieval Knowledge Source

Where each trial gets context

Corpus Directory

Required local evidence to parse, chunk, embed, and index.

Web Search Only

Skip local indexing and answer from live web results.

Adaptive RAG: Corpus + Web

Use the local corpus first, then fall back to web search when context is weak.

tune Pipeline Mode help

Full RAG Generate an answer per trial; score retrieval and/or generation metrics.

Retrieval-Only Test retrieval quality only -- no answer generated, no main LLM called.

settings_suggest Answer Generation LLM

Generates the final answer from local or web context

Retrieval Corpus Path * Corpus used for indexing and evaluation-data generation.

Answer LLM Provider *

Answer LLM Model *

API Key Never saved to disk

Base URL (Optional)help

dataset Corpus & Evaluation Data

Required inputs for the selected knowledge source

document_scanner Document Parser (OCR)

Optional — scanned / complex documents

Parserhelp

Use OCR for indexing help

payments

Custom Pricing

Optional per-run LLM token rates. Default catalog pricing remains active for models you do not override.

Default catalog pricing

Override default LLM pricing for this run Enter USD per 1 million input and output tokens. Both rates are required; zero is valid.

dataset Evaluation Metrics: 0 active

verified Scored against ground-truth Q&A pairs across retrieval & generation quality

info

Metrics choose what is scored, not which pipeline stages execute. Full RAG always retrieves and generates an answer for every sample; generation metrics add judge-model scoring. Retrieval-Only skips answer generation.

01 Dataset Source Strategy help

auto-generate

Ground-truth Q&A evaluation dataset used to objectively score candidate RAG architectures

Auto-generate on launch Default

Composer creates evaluation Q&A pairs from the selected retrieval corpus at launch.

Use an existing dataset file

Supply ground-truth QA triplets for scoring. This does not replace the retrieval corpus.

02 Validation Split Policy help

Full Set (No Split)

Hold out an unseen test split for scoring to prevent configuration overfitting and selection bias

Split into train/test, score on test only

Disabled in Auto Mode

Test Split Proportion help Held-out fraction reserved for objective trial scoring

Reserved: 80% (not scored) Test: 20% (scored)

03 Retrieval Metrics help

0 active

Candidate passage relevance & ranking accuracy. These metrics add no judge-model calls; Full RAG answer generation still runs.

speed Instant deterministic evaluation across document contexts

04 Generation Metrics (LLM Judge) help

0 active

Deep semantic answer correctness and faithfulness. Each selected metric is judged independently and can add cost.

Use a different judge LLM help

psychology Each selected judge metric can add one LLM call per sample per trial

speed Run Limits

Cost, concurrency, and compute

Samples per Trialhelp

Parallel Trialshelp

Maximum Trials (Optional)help

Embedding / Reranker Devicehelp

tune Total Search Space Combinations: 0

info Combinations are computed across all active stages below

01 Query Expansion

2 active

Multi-query fanout, HyDE hypothetical docs, or direct query. Select "none" too if you want the search to compare with vs. without expansion.

02 Retrieval Strategy

2 active

Candidate vector similarity, MMR diversity, or hybrid BM25 + dense

03 Reranking & Fusion

1 active

Choose built-in strategies or connect your own reranking service. Select “None” too to compare with and without reranking.

04 Retrieval Depth (k)

2 active

Number of top candidate context passages passed to LLM

05 Chunking Strategy

0 active

Pick an algorithm, set size/overlap, then Add — repeatable

Method

Size

Overlap

06 Embedding Models

0 active

Dense embedding models — Arabic and English supported, type any HuggingFace model id

07 Vector DB Backends

0 active

Target index storage & vector search engine backends

edit_note

Prompt Manager

Only prompts executable by this run are shown. Changes apply to every trial and never edit packaged defaults.

Prompt language

privacy_tip Do not paste credentials or secrets into prompts. Required variables must remain present so runtime formatting stays safe.

fact_check Review Before Launch

Complete submitted configuration; sensitive keys are never displayed

edit_note Resolved Prompts

Arabic · defaults

account_tree Search Space Configuration

Every selected value by stage

view_list Resolved Search Space Combinations

0 combinations

These are the exact pipeline trials that will run before an optional maximum-trials cap is applied.

verified_user

Validate before launch

Run the dry-run check to verify the current providers, dataset, metrics, and search space. Launch appears only after it succeeds.

lock Changing any setting requires a new dry run