Search Architecture
Four-Stage Hybrid Retrieval
Every query passes through a cascading pipeline that combines lexical precision with semantic understanding. Each stage narrows and reranks candidates until only the most relevant grants remain.
67,000+
grants indexed
12
data sources
<200ms
scoring latency
$0.003
per query
<1ms
query time
Full-Text Search
PostgreSQL ts_rank over indexed grant titles, descriptions, and eligibility criteria. Sub-millisecond candidate retrieval from 67,000+ opportunities.
3×
recall lift
Query Expansion
LLM-generated synonym sets and domain-specific rewrites broaden recall without sacrificing precision. A query for "youth STEM" also surfaces "K-12 science education" and "after-school technology programs."
768-d
vectors
Embedding kNN
Dense vector search with pgvector finds semantically similar grants that keyword matching misses. Captures conceptual overlap across different funding vocabularies.
60.3%
P@5
Cross-Encoder Reranking
A fine-tuned cross-encoder jointly attends to the full query–document pair, producing calibrated relevance scores that outperform bi-encoder similarity alone.
Federated Search
Five Providers, Five Web Indices
For open-web grant discovery, we query five LLM providers simultaneously—each backed by a distinct search index. Results are fused under a single learned scoring function that ignores every model's self-reported relevance score.
Gemini
Google Search index
GPT-4.1
Bing/OpenAI index
Claude
Anthropic web search
Grok
X/Twitter + web index
Perplexity
Independent web index
Key Insight
We discard every model's self-reported confidence score and re-evaluate all candidates through our own 15-feature scoring function—trained on 1,034 labeled query–grant pairs.
Scoring Model
15-Feature Learned Relevance
A gradient-boosted model scores every candidate grant across five feature categories. Trained on 1,034 labeled queries, validated at 60.3% Precision@5—then distilled into a lightweight scorer for sub-200ms inference.
Text
- BM25
- TF-IDF overlap
- Title match
Semantic
- Cosine similarity
- Cross-encoder score
- Query expansion hits
Metadata
- Agency match
- Category alignment
- Eligibility fit
Freshness
- Days to deadline
- Posted recency
- Update frequency
Penalty
- Expired flag
- Duplicate detection
- Low-quality signals
1,034
labeled queries
15
scoring features
60.3%
Precision@5
~0ms
distilled inference
Data Pipeline
12 Sources, Real-Time Ingestion
We ingest grant data from 12 federal and institutional sources, normalize schemas, deduplicate listings, and maintain freshness with daily sync jobs. 99.98% of opportunities are fully tagged with eligibility, category, and deadline metadata.
67,000+
opportunities
99.98%
fully tagged
Daily
sync cadence
Grant Writing Engine
Six Steps from RFP to Polished Draft
Every grant has unique requirements. Granted's workflow ensures each one is identified, addressed, and woven into a draft that speaks directly to your funder.
RFP Analysis
Upload your RFP or grant guidelines. Granted’s AI reads the full document and identifies every required section, evaluation criterion, and compliance requirement.
Requirement Discovery
The system discovers the grant’s full structure—from project narratives and budget justifications to data management plans and letters of support.
Grant Writing Coach Q&A
A grant writing coach asks targeted questions about your organization, team qualifications, project goals, and budget. Your answers ground every section in your real data.
Coverage Tracking
Track which requirements have been addressed and which need attention. See coverage percentage in real time as the coach gathers information.
Section-by-Section Drafting
Each section is drafted individually using your specific answers and the RFP’s requirements. No generic templates, no placeholders.
Purpose-Built, Not General-Purpose
General-purpose AI doesn’t read your RFP, track coverage, or ground output in your data. Granted does—because it was built for this one job.
Read the Technical Paper
Full methodology, ablation studies, and benchmark results for the hybrid retrieval pipeline and knowledge-distilled scoring model.
What Makes This Different
ChatGPT, Claude, and other general-purpose AI tools are powerful writers—but they weren't designed for grant proposals. Here's what Granted does that they don't.
Your Data Stays Yours
Everything you upload to Granted—your RFP, your coach answers, your drafts—is private to your account. We never use your data to train models, and we never share it with third parties.