Brainstorm Search Engine

• Brainstorm Search Engine is an information retrieval system that surfaces diverse, verifiable stepwise derivations from a curated scientific knowledge base.
• It employs large language models for generating and verifying long chains-of-thought (LCoTs) using a hybrid of keyword indices and semantic embeddings.
• The system enhances research by using cross-model consensus and diversity-aware ranking to minimize factual errors and promote innovative scientific synthesis.

A Brainstorm Search Engine is an information retrieval system designed to surface diverse, deeply reasoned, and cross-domain connections among scientific knowledge points by tracing explicit derivational chains (Long Chains-of-Thought, LCoTs). In contrast to traditional retrieval methods that compress or abstract away underlying reasoning, the Brainstorm paradigm targets inverse knowledge search: given an endpoint concept, retrieve distinct, verifiable stepwise derivations from a curated knowledge base. This approach underpins scalable scientific synthesis, improves factual fidelity, and fosters discovery of novel associations across disciplines (Li et al., 30 Oct 2025).

1. Formalization of Inverse Knowledge Search

Let $K = \{C_1, C_2, \ldots, C_n\}$ denote the set of fine-grained scientific knowledge points (e.g., phenomena, theorems, mechanisms), each referenced by canonical textual identifiers. The central knowledge base $B = \{(Q_i, A_i, \mathrm{LCoT}_i)\}$ consists of triplets:

• $Q_i$: a rigorously formulated first-principles question (endpoint),
• $A_i$: a verifiable answer,
• $\mathrm{LCoT}_i$: an ordered list of deductive steps from foundational premises to $A_i$.

The inverse knowledge search problem is: given a target concept $C_t \in K$, retrieve a ranked set $S \subset B$ of LCoT derivations satisfying

• Coverage: $C_t$ explicitly appears within steps of $\mathrm{LCoT}_i$,
• Diversity: $B = \{(Q_i, A_i, \mathrm{LCoT}_i)\}$0 contains derivations spanning multiple scientific domains and abstraction levels.

Formally, the retrieval function $B = \{(Q_i, A_i, \mathrm{LCoT}_i)\}$1 maps $B = \{(Q_i, A_i, \mathrm{LCoT}_i)\}$2, maximizing (i) recall of valid derivations referencing $B = \{(Q_i, A_i, \mathrm{LCoT}_i)\}$3 and (ii) cross-domain diversity among chains (Li et al., 30 Oct 2025).

2. System Architecture and Workflow

The Brainstorm Search Engine operates within a five-stage pipeline:

• A. Data Ingestion
  • A curriculum-structured coverage (∼200 courses × ∼200 topics) defines endpoints.
  • LLM planners and generators produce ∼3 million rigorously posed questions.
• B. LCoT Generation & Verification
  • Multiple independent solver LLMs each construct stepwise derivations for every $B = \{(Q_i, A_i, \mathrm{LCoT}_i)\}$4.
  • Chains are filtered by prompt sanitization and retained only if solvers' answers concur (cross-model consensus).
• C. Indexing
  • Keyword and embedding indices are created: terms and chains’ semantic representations (using embedding function $B = \{(Q_i, A_i, \mathrm{LCoT}_i)\}$5) are stored for subsequent fast retrieval.
• D. Retrieval (Brainstorm Engine)
  • User queries are expanded to a set of keywords $B = \{(Q_i, A_i, \mathrm{LCoT}_i)\}$6 and a semantic embedding $B = \{(Q_i, A_i, \mathrm{LCoT}_i)\}$7.
  • Candidate chains are jointly generated via exact keyword match and Approximate Nearest Neighbor (ANN) search in embedding space.
  • Composite relevance/dissimilarity metrics rank and filter the candidates.
• E. Consensus Verification
  • Candidate chains are re-verified by cross-model majority vote on answer consistency and re-sanitized for prompt drift.
  • Only chains passing full end-to-end validation are retained (Li et al., 30 Oct 2025).

This pipeline guarantees that every retrieved derivation is (i) explicable in first principles, (ii) independently verified, and (iii) efficiently accessible by both lexical and semantic criteria.

3. LCoT Knowledge Base Representation and Linkage

Each knowledge base entry $B = \{(Q_i, A_i, \mathrm{LCoT}_i)\}$8 is stored with strong normalization:

• $B = \{(Q_i, A_i, \mathrm{LCoT}_i)\}$9: plain text with metadata (course, topic, prompt type),
• $Q_i$0: structured numeric, symbolic, or categorical answer,
• $Q_i$1 : a sequence of steps $Q_i$2, each as independently embeddable text (optionally with code/math).

Inverted indices map every concept $Q_i$3 to the set of chain IDs where $Q_i$4 appears in any step; semantic embeddings are computed for both questions and individual steps. Chains are also grouped by topic and abstraction level, supporting layered and diversity-aware retrieval (Li et al., 30 Oct 2025).

4. Retrieval and Ranking Algorithm

Brainstorm retrieval proceeds as follows:

• Step 1: Expand the query $Q_i$5 to include synonyms/related terms; compute embedding $Q_i$6.
• Step 2: Generate candidates via union of keyword index ($Q_i$7 and synonyms) and embedded ANN search (top-$Q_i$8 chains by cosine similarity $Q_i$9, where $A_i$0).
• Step 3: Score each candidate $A_i$1 with

$A_i$2

where $A_i$3 are hyperparameters (e.g., $A_i$4), $A_i$5 denotes frequency-normalized presence of $A_i$6 in $A_i$7, and $A_i$8 penalizes near-duplicates.

• Step 4: Discard candidates below threshold $A_i$9 and enforce a domain-quota to ensure cross-disciplinary breadth.
• Step 5: Cross-model consensus verification: for each remaining chain, solvers regenerate answers, and the original $\mathrm{LCoT}_i$0 is retained only if it matches the majority result; logical consistency is re-checked via independent verification.

Pseudocode is provided in the source (Li et al., 30 Oct 2025), encoding all retrieval and consensus logic.

5. Quantitative Evaluation

In controlled evaluations across 200 topics and six scientific domains, Brainstorm-powered synthesis (via the Plato agent) yields:

Metric	Plato (Brainstorm LCoTs)	Baseline (LLM, no retrieval)
Knowledge-point density	+45% over baseline	Reference value
Factual error rate	6% per 1k words	12% per 1k words
Median domain diversity	0.78	0.45 (random retrieval)

Knowledge-point density is the distinct, verifiable concepts per article. Factual error rate is judged by an external LLM (GPT-5). Diversity is measured as average pairwise domain distance among chains for a target concept (Li et al., 30 Oct 2025).

These results demonstrate that inverse knowledge search over LCoTs enables higher-density, lower-error, and more cross-disciplinary scientific synthesis compared to direct generative approaches.

6. Integration with Synthesis and Practical Implications

The Brainstorm engine underlies the SciencePedia encyclopedia generation pipeline: for a user-specified endpoint $\mathrm{LCoT}_i$1, it retrieves and verifies chains $\mathrm{LCoT}_i$2, which are then synthesized into encyclopedic entries by Plato, a downstream LLM, following domain-specific style guides. This structure:

• Grounds articles in explicit, stepwise reasoning, preventing content drift and hallucination.
• Scaffolds the synthesis process, enhancing narrative coherence.
• Injects cross-disciplinary breadth beyond surface keyword or nearest-neighbor retrieval.

Operational limitations include increased retrieval latency for very high-frequency endpoints, cold-start lag for newly introduced concepts, and intrinsic dependence on the temporal coverage of LLM training cutoffs. Mitigating steps involve clustering for speed, on-demand Socratic generation for coverage, and plans for formalizing LCoT structure into graph databases and extending to continuously updated sources such as textbooks and research articles (Li et al., 30 Oct 2025).

7. Distinctiveness and Relation to Other Brainstorm Engines

The Brainstorm architecture departs sharply from analogy engines (Kang et al., 2022), which match on high-level purpose or mechanism embeddings to foster creative ideation through analogical transfer, and from meta-search graph engines using platforms like Wikipedia for semantic exploration (Fuehres et al., 2012). Unlike those, Brainstorm’s verifiable, stepwise, endpoint-anchored LCoT retrieval provides not only domain-overlapping context but also substantiates inferences with explicit derivational evidence traceable through logical steps. This reasoning-centric approach is essential for both scientific verification and scalable synthesis (Li et al., 30 Oct 2025).