Search that uses engineered knowledge, not raw text alone
Queries can be matched against canonical symptom descriptors, multilingual captions, crop metadata, and source-linked text instead of relying only on keyword overlap.
FieldKB , the evidence layer behind every Field State answer
A Field State answer rests on two things: your own farm's records, and grounded agronomy knowledge. FieldKB is the layer that keeps both source-traceable, so a farmer can ask why, and check it: why this number, why this guidance, and whether it came from their own records or from verified knowledge.
Multilingual search inputs and outputs
Crop × country × farm-memory knowledge layers
Verified, farm-specific, and local-practice evidence
Every answer traces back to a source
Your own records
How Field State answers from your own farm
Grounded agronomy knowledge shapes general guidance. The bigger half is your farm's own record: Field State answers from what you have actually logged. Three layers make that trustworthy.
01 · Load
It keeps your records, with proof
Each record holds what happened, when, who or what it involved, and how it is known: a photo, a receipt, or an honest "reported only". It is saved on your phone first, so it holds even with no signal, and is remembered cycle after cycle.
02 · Retrieve
It answers from them, with the math shown
Ask a plain question and the answer is worked out from your own records by exact calculation, with a receipt of the exact records behind every figure. It answers only from your farm's history; it does not invent it.
03 · Extract
You speak, it proposes, you confirm
A spoken note becomes a structured record. The AI proposes the pieces, the amount, the quantity, who it was with, what it was for, and you confirm or fix them. Only what you confirm counts toward your numbers.
The AI proposes and explains; it never decides a figure or turns an unconfirmed guess into a fact. Money is added up by exact calculation, and every total opens to the records behind it.
Research framing
More than search: structured feature engineering for agriculture
FieldKB is not just a document index and not just an image gallery. It combines domain-informed curation, qualitative pattern analysis, canonical symptom design, and multilingual metadata construction to make agricultural retrieval more precise and easier to audit.
- We review recurring visual and textual patterns across crops, issues, and growth stages.
- We normalize those observations into canonical descriptors so search does not drift across synonyms.
- We design metadata for discrimination, so visually similar diseases and pest damage can be separated more reliably.
- We link answers back to human-readable material such as curated notes, image descriptors, and source-aware references.
- We keep scope and uncertainty visible, so weak evidence can be handled differently from strong evidence.
Pattern analysis
Canonical symptom modeling
Retrieval-oriented feature engineering
Multilingual knowledge design
How knowledge flows through FieldKB
Step 01 · Curate
Collect agronomy material, field notes, and images
We organize documents, practice notes, regulations, and field images into crop- and country-aware knowledge layers with explicit source handling and structured metadata.
Step 02 · Analyze
Extract patterns and normalize features
We identify recurring symptom patterns, lesion morphology, scene cues, and contextual signals, then normalize them into canonical tokens and retrieval-friendly descriptors.
Step 03 · Structure
Build multilingual, retrieval-ready records
Features are serialized into structured records that support semantic search, filtering, multilingual retrieval, and clearer separation between similar diseases, pests, and stress patterns.
Step 04 · Retrieve
Return evidence-linked answers
Search results are assembled from the structured knowledge layer so users can inspect not only the answer, but also the supporting descriptors, examples, and source-aware context behind it.
Core capabilities
What FieldKB is built to support
The goal is not to replace agronomists or local judgment. The goal is to make agricultural knowledge easier to search, compare, review, and explain.
Retrieval across documents, symptoms, and field imagery
FieldKB is designed to connect practice guidance, issue descriptions, and example images, so users can compare written evidence with visual patterns in one workflow.
Crop, country, and deployment context matter
A useful agricultural answer depends on crop, geography, local practice, and knowledge coverage. FieldKB is structured to route retrieval through those layers instead of pretending every answer is universal.
Research rigor
How rigor is built into the system
FieldKB is designed to make the knowledge layer explainable: what was observed, how it was normalized, and how it is later used for retrieval.
- Controlled vocabularies reduce annotation drift across crops, diseases, pests, and image sets.
- Symptom descriptors are chosen for discriminative value, not just descriptive completeness.
- Multilingual records are designed to preserve retrieval meaning across languages, not just literal translation.
- The schema separates storage, visual context, search controls, localization, and diagnostic linkage for cleaner auditing.
4
Research pillars: curation, analysis, normalization, retrieval
2+
Retrieval modalities: text and image-linked metadata
1
Structured knowledge contract behind each indexed asset
Values above describe design intent and system structure. They are not claims about benchmark performance, index size, or universal coverage.
Data and provenance
Data handling, provenance, and licensing
FieldKB is built around source-aware organization. Public descriptions remain high level here; detailed internal schemas, validation rules, and pipelines are intentionally not exposed.
What material goes into FieldKB?
- Agronomy notes, crop issue descriptions, field observations, and image metadata prepared for retrieval.
- Public or partner material handled according to usage rights, access boundaries, and deployment scope.
- Curated multilingual descriptors that make search and downstream review more consistent.
How licensing and source boundaries are treated
- Source handling is tracked separately from retrieval metadata so rights information does not pollute search embeddings.
- Restricted or partner-specific content can be isolated into scoped layers rather than mixed into general access.
- What users see in the interface is a human-readable, deployment-appropriate view rather than raw internal identifiers.
Provenance and updates
- Records can be revised as guidance, regulations, or evidence quality changes over time.
- Region- and crop-specific layers make it possible to update one deployment scope without changing unrelated ones.
- FieldKB organizes and retrieves information; it does not replace the original issuing authority for official documents.
FieldKB is a structured retrieval and decision-support layer. Official labels, regulations, and partner documents remain authoritative in their original form.
Boundaries and responsible use
What FieldKB does, and what it does not claim to do
Strong systems are honest about scope. FieldKB is designed to improve retrieval quality and evidence handling, but it is still a support layer that depends on knowledge coverage, review quality, and deployment context.
Scope of answers
- Focused on agricultural retrieval within the crops, issues, regions, and knowledge layers that have been curated.
- Not a replacement for local agronomists, extension teams, legal review, or stewardship procedures.
- Some queries may correctly return limited results or no answer when evidence is weak or outside current scope.
Privacy and deployment boundaries
- Private or partner material can be segmented into restricted layers depending on deployment needs.
- Field observations used for improvement should be reviewed for privacy, consent, and organizational policy.
- The system is not intended as a store of unrelated personal or sensitive non-agricultural data.
Responsible use in practice
- Treat outputs as decision-support that should be interpreted alongside local expertise and field context.
- When retrieved guidance conflicts with an official label or regulation, the official source takes priority.
- New crops, countries, or deployments should be reviewed for knowledge coverage before they are presented as validated for use.
Internal ranking methods, embeddings, and infrastructure choices are intentionally not detailed here. What matters for users is that the system is structured, reviewable, and explicit about its boundaries.
Interested in FieldKB research, pilots, or knowledge integration?
We work with research teams, agricultural organizations, and deployment partners who need structured, multilingual, evidence-linked agricultural knowledge systems.