Keyword plus embeddings
Exact matches and embedding neighbors share a score. Recency and field weights sit in the same query.
Product
Discovery infrastructure, whole team.
Gleanor is ranking infrastructure: hybrid catalog search, cited RAG, and recommendations on one index. Hosted cloud, or a license that stays on your floor.
Cited RAG
Excerpts with source ids, packed for a generator. Filters on the ticket stay on the pack that comes back.
Recommend
Item-to-item and user-to-item from the same index. A catalog change is a related-item change.
Studio
Operators tune ranking. The API reads the same published profile. People publish; nothing runs a desk on its own.
AI in the product
Three model jobs. One index.
Each use case names the model job, who it is for, and the route. Gleanor is a core AI/ML retrieval product. It is not an agent runtime.
Use case · Hybrid ranking
Embeddings join the keyword score
For search engineers and merchandisers
Keyword hits and dense neighbors share a fused score. Cohere Embed is the planned dense encoder. Recency and field weights apply after the fuse, so Studio can bias a season without a new embedding job.
Workflow: connect source → embed + invert → Studio weights → POST /v1/search
Use case · Cited RAG
Retrieved chunks, sized for a generator
For support leads and help-center teams
Gleanor retrieves and packs cited windows. The generator is yours: GPT-4o, Llama 3 70B, or AWS Nova are the planned options. Filters on the ticket survive into the pack. We do not train that generator, and we do not run an agent over it.
Workflow: query + filters → retrieve → cited window → POST /v1/rag → your model
Use case · Recommendations
Neighbors from the same embedding space
For merchandisers and wiki owners
Item-to-item and user-to-item read the same corpus id as search. A catalog change is a related-item change. No second pipeline, no separate embedding store to drift.
Workflow: item or user id → neighbor search → POST /v1/recommend
Not an agent
Studio is a human publish step
For operators who own relevance
People draft a ranking profile, preview it on a saved query set, then publish. The API pin is that version. Gleanor is a retrieval tool — not an agent framework.
Workflow: draft → preview queries → publish → production pin
Planned model access: Bedrock, OpenAI direct, and Hugging Face. Exact encoder and generator are chosen on the first corpus call — we do not claim a live multi-model roster today.
Index
What a published profile carries
Ranking fuse
Keyword hits and embedding neighbors share a score. Recency and field boosts apply after the fuse, so a merchandiser can bias without a new embedding job.
Sourced windows
RAG returns excerpts with a source id and a span. Filters on the query survive into the pack — locale, brand, and collection stay attached.
Recommendations
Item-to-item and user-to-item read the same corpus. Related SKUs and related runbooks do not live on a second pipeline.
Studio publish
Operators draft a profile, preview on a saved query set, then publish. The API pin is the published version.
Sources
Sources that already fit
Catalogs
Product feeds and merchandising tags. Studio holds seasonal boosts without a re-embed of the whole barn.
Help centers
Crawl or webhook. RAG answers keep the filter the ticket already set, and every excerpt carries a source id.
Internal wikis
S3 drops and CSV. Recommend surfaces related runbooks from the same corpus id.