The brief
Bodh is an offline semantic knowledge retrieval system designed to give organisations a private and customisable way to search their internal technical knowledge.
The system was originally built around knowledge covering computer-science tools, technical documentation, company-specific usage, internal practices, and other custom knowledge points.
Traditional search performs well when users know the exact words they are looking for. Technical knowledge rarely works that way.
A developer may describe a problem instead of naming the tool that solves it, use different terminology from the documentation, or ask a natural-language question whose answer is conceptually related to a document without sharing the same keywords.
Bodh was designed to retrieve information based on meaning rather than exact wording.
A private knowledge layer
Organisations can build their own knowledge collection containing documentation, technical references, internal practices, tool information, usage guidelines, and company-specific knowledge.
The retrieval system can operate locally without depending on an external cloud AI service.
This allows the organisation to control both the information stored in the system and the environment in which queries are processed.
The result is a private knowledge layer that can be customised around the organisation rather than a generic external knowledge service.
Search by meaning
Keyword search relies heavily on vocabulary.
A document might discuss container orchestration while a developer asks how the company manages Docker deployments. Although the two pieces of text may use different words, they can refer to closely related concepts.
Bodh approaches retrieval semantically.
Stored knowledge is transformed into a representation that captures characteristics of its meaning. User queries are transformed into the same representation space.
The system can then search for knowledge that is conceptually close to the query rather than requiring an exact textual match.
Building the knowledge index
Before retrieval begins, the available knowledge is processed into suitable knowledge units.
These units are transformed into compact representations and stored in an index.
This separates knowledge ingestion from query-time retrieval.
The ingestion pipeline follows:
Knowledge Sources → Processing → Knowledge Units → Latent Representation → Search Index
Once the index has been created, the system can efficiently search the stored knowledge without repeatedly processing the entire corpus.
Query-time retrieval
When a user asks a question, Bodh transforms the query into the same representation space used by the indexed knowledge.
The query representation is compared against the stored representations, allowing the system to identify the knowledge points that are semantically closest to the user’s request.
The retrieval path becomes:
User Question → Query Representation → Indexed Similarity Search → Relevant Knowledge
This creates a natural-language search experience over organisation-specific information.
Learning a latent knowledge space
One of the more experimental parts of Bodh was its representation-learning architecture.
The system used an autoencoder-based approach and split the representation pipeline around the model’s latent space.
Knowledge could be encoded into compact latent representations and indexed for retrieval.
Queries were processed into compatible representations and compared against the stored knowledge.
This allowed the system to explore semantic retrieval using learned representations rather than depending exclusively on lexical search.
Custom organisational knowledge
Bodh was designed to work with knowledge that matters specifically to an organisation.
A company could build a knowledge collection around:
- Internal development tools
- Technology documentation
- Company-specific tool usage
- Engineering conventions
- Infrastructure practices
- Deployment procedures
- Internal workflows
- Technical references
- Custom operational knowledge
The retrieval layer therefore adapts to the organisation’s knowledge rather than requiring the organisation to rely on a fixed public corpus.
Offline by design
Bodh can operate without sending organisational knowledge or user queries to an external AI provider.
This provides several useful properties.
Privacy — internal knowledge can remain within the organisation’s environment.
Control — the organisation controls the knowledge base and retrieval infrastructure.
Availability — retrieval can continue without an internet connection.
Customisation — the corpus can be built entirely around organisation-specific information.
This makes the architecture useful for environments where sensitive technical documentation cannot be freely transmitted to external services.
Before modern RAG workflows
Bodh explored semantic organisational knowledge retrieval before retrieval-augmented generation became a mainstream application pattern.
Its core workflow already followed an important part of the modern retrieval architecture:
Store organisational knowledge → Build semantic representations → Index knowledge → Understand a query → Retrieve relevant context
Bodh focused specifically on the retrieval layer rather than requiring a generative model to provide the intelligence.
This makes the system useful independently and also provides a foundation upon which a conversational or generative layer could be added.
A foundation for private AI
A reliable retrieval system is one of the core building blocks of a private knowledge assistant.
Once relevant organisation-specific information can be identified efficiently, another intelligence layer can use that context to answer questions, explain internal systems, assist developers, or support operational workflows.
Bodh therefore separates organisational knowledge from any particular AI provider or model.
The organisation owns the knowledge and controls how it is retrieved.
The engineering idea
The project was built around a simple principle:
Your organisation should be able to understand and retrieve its own knowledge without giving that knowledge away.
By combining semantic indexing, learned latent representations, offline retrieval, and custom organisational knowledge, Bodh turns technical documentation into a private searchable knowledge layer.
Project highlights
- Fully offline semantic knowledge retrieval
- Organisation-specific knowledge bases
- Natural-language querying
- Semantic rather than purely lexical matching
- Autoencoder-based representation learning
- Latent-space indexing
- Query representation and similarity search
- Private local knowledge storage
- Custom technical documentation support
- No dependency on cloud AI services for retrieval
- Foundation for private knowledge assistants and RAG-style systems