The brief

Most organisations already have the information their teams need — stored in technical documentation, internal guides, development practices, infrastructure notes, tool references, and operational knowledge. The hard part is finding the right piece when the person searching doesn’t know the exact document, phrase, or keyword used to describe it.

Solving that without sending the organisation’s knowledge to an external AI service was the brief behind Bodh. It creates a private, offline search layer that understands queries by meaning rather than relying only on exact word matches. Teams can build a knowledge base around their own tools, systems, terminology, and working practices, then search it naturally from within their own environment.

Why Bodh works the way it does

A simple observation was the starting point: internal knowledge is often difficult to use even when it’s well documented.

A developer might ask:

“How do we deploy this service with the tools used by our team?”

But the documentation might describe the same process using terms such as container orchestration, deployment manifests, or a specific internal platform name.

Because the question and the document don’t share the same wording, a keyword search can fail outright. Someone familiar with the organisation would understand the connection immediately — and providing that kind of meaning-based connection through a private retrieval system is what Bodh was designed for.

Building a generic chatbot was never the goal. Making an organisation’s own knowledge easier to find, while keeping control of the data and the environment where search takes place, was.

Search by meaning, not exact wording

When users know the exact terms used in the source material, traditional search performs well.

Semantic search takes a different approach: it represents both the user’s query and the stored knowledge as numerical embeddings, then looks for representations close in meaning. Related concepts, alternative phrasing, and natural-language questions all become easier to handle than they would be with exact keyword matching alone.

A user might search for:

“What is our process for rolling back a failed release?”

While the relevant document is titled:

“Production deployment recovery procedure”

Different wording, closely related intent — identifying that relationship is exactly what Bodh is built to do.

Creating the knowledge base

Before anyone can search, the organisation’s knowledge needs preparing. Documents and other knowledge sources get divided into useful knowledge units. Each unit is transformed into a compact representation and stored in a searchable index.

The ingestion process follows:

Knowledge sources → Processing → Knowledge units → Learned representations → Search index

Once the index exists, queries can be compared against the stored representations without repeatedly processing the entire knowledge collection — keeping ingestion and retrieval separate, and making the system easier to update as new documentation and internal practices are added.

Learning a representation of organisational knowledge

An autoencoder-based representation-learning approach set Bodh apart from a typical search tool.

The project explored how technical knowledge could be encoded into a learned latent space — a compact representation where related concepts sit closer together. Documents were encoded into this space and indexed for retrieval, while user queries were processed into compatible representations so the system could compare them against the indexed knowledge.

The workflow was:

Knowledge text → Encoder → Latent representation → Similarity index

At query time:

User question → Query encoder → Latent representation → Similarity search → Relevant knowledge

This let the project investigate semantic retrieval through learned representations rather than relying only on word overlap.

Knowledge built around the organisation

Knowledge specific to the organisation using it is what Bodh was designed for. A company could build a private collection around internal development tools, engineering conventions, infrastructure practices, deployment procedures, technology documentation, company-specific terminology, tool usage guidelines, operational workflows, and technical references.

That makes the system more useful than a generic public search tool for questions depending on local context. Finding a document isn’t the whole value — finding the document that reflects how this particular organisation works is.

Offline by design

Entirely within a local or private environment, Bodh can run. The organisation controls the knowledge collection, the retrieval process, and the infrastructure used to answer queries — internal documentation and user questions never need to be sent to a cloud AI provider for the retrieval step.

Several practical benefits follow from this:

  • Sensitive technical information can remain inside the organisation’s environment.
  • Search remains available without an internet connection.
  • The knowledge base can be customised around internal terminology.
  • The organisation retains control over indexing and retrieval.
  • The system can be used in environments with strict data-handling requirements.

For private technical knowledge, offline capability isn’t just a deployment preference. It can be part of the product’s trust model.

Retrieval without generation

The retrieval layer itself is where Bodh focused its effort. It didn’t need a generative model to produce a conversational response — its primary job was identifying the most relevant pieces of organisation-specific knowledge for a given question.

That makes the system useful on its own, and also gives it a clear role in a larger private-AI architecture:

Private knowledge → Semantic retrieval → Relevant context → Optional assistant or generation layer

Because of this separation, an organisation can improve or replace the downstream answer-generation layer without giving up ownership of the knowledge index and retrieval system.

A foundation for private assistants

Reliable retrieval is what makes a knowledge assistant useful in the first place. Bodh provides one of the core building blocks for a private assistant: locating relevant internal knowledge based on what a user means, not only on the exact words they typed.

Developer support, internal documentation search, infrastructure assistance, onboarding and training, technical troubleshooting, internal process discovery, and private RAG-based assistants could all sit on top of that foundation. The retrieval system remains independent of any single external model provider.

Outcome

Organisation-specific technical knowledge became a searchable, offline semantic index — that’s what Bodh delivered. The client could search internal documentation and practices using natural-language questions, even when the query didn’t share terminology with the source material.

The project delivered:

  • A fully offline retrieval capability.
  • A private, organisation-controlled knowledge layer.
  • Meaning-based search through learned representations.
  • Autoencoder-based latent-space indexing.
  • A foundation for future private knowledge assistants.

What came out of it was a retrieval system built around a clear principle: useful organisational knowledge should remain both accessible to the people who need it and controlled by the organisation that owns it.

Project highlights

  • Built an offline semantic knowledge-retrieval system in 7 weeks.
  • Enabled private, organisation-controlled knowledge bases.
  • Supported natural-language queries over technical documentation and internal practices.
  • Used embeddings and learned latent representations for meaning-based retrieval.
  • Implemented an autoencoder-based representation-learning pipeline.
  • Indexed knowledge units for similarity search.
  • Supported custom organisational terminology and tool knowledge.
  • Avoided dependence on cloud AI services for retrieval.
  • Separated knowledge ingestion from query-time search.
  • Created a foundation for private assistants and RAG-style applications.
  • Designed the system for sensitive technical and operational knowledge.