gita / reflection Find a passage
/ DUTY & ACTION/ CLARITY OF MIND/ EVERYDAY LIFE

Bring a real situation.
Discover a perspective from
the Bhagavad Gita.

700 verses. One grounded answer.

Bhagavad
Gita.

Scroll to explore the dialogue

Wisdom for your moment

A question from your life.
A response grounded in
cited verses.

Pause.
Find a
perspective.

Describe what you are facing. Explore a practical reflection alongside the original Sanskrit and its English translation.

01

Begin with your situationWork, attention, anger, or a difficult decision.

02

Find supporting versesSearch all 700 verses of the Gita.

03

Reflect with contextRead the response alongside its cited sources.

Generate a grounded answer

0 / 2000

Try an example

⌘/Ctrl + Enter to ask

Grounded reflection

Reflection

Original Sanskrit

Transliteration

English translation

Citation grounding

Cited verses

Claim support

Latency for this response

Query generation
Retrieval
Reranking
Selection
Final answer generation
Total pipeline
How this answer was found

Queries sent to retrieval


            

            

              
RRF fused candidates

              

            

            

              
Claude evidence ranking

              

            

            

              
Selected grounding verses

              

            

          

        

      


      

        

          700
          
Verses cached locally

          

            Every Bhagavad Gita verse keeps its translation, metadata, and
            precomputed vector ready in memory for fast search.
          

        

        

          4
          
Retrieval facets generated

          

            Claude generates four focused facets. The original question and
            deterministic Gita-specific expansions are searched as well.
          

        

        

          20
          
Candidates reranked

          

            Claude ranks up to 20 fused candidates against the natural question
            and returns the six strongest evidence candidates.
          

        

      


      

        How each answer is grounded 
      

        

          

            
Every stage of the RAG pipeline

          

          

            The production path combines Amazon Bedrock, a local vector cache,
            deterministic retrieval controls, RRF fusion, and Claude evidence reranking. Each
            stage does one narrow job well, from translating the question into
            search concepts to validating the final evidence-backed answer.
          

        


        

          

            
Corpus preparation

            

              The preparation job validates and enriches 700 verse records, then
              persists translations, metadata, themes, and embeddings for retrieval.
            

            

              load_records()
              enrich()
              validate()
            

          


          

            
Query intelligence

            

              Claude Sonnet creates four retrieval facets, including ideal evidence,
              while a controlled domain map adds Gita-specific concepts and synonyms.
            

            

              generate_search_queries()
              expand_domain_concepts()
              Claude Sonnet 4.6 (default)
            

          


          

            
Hybrid retrieval

            

              Titan semantic search and BM25 lexical search run across the local
              700-verse cache for the original question, generated facets, and
              deterministic domain expansions.
            

            

              retrieve_candidates()
              Amazon Titan Text Embeddings V2
              local_vector_search.py
            

          


          

            
Rank fusion

            

              Reciprocal Rank Fusion merges results from every search, rewards
              repeated evidence, removes duplicate verses, and keeps the top 20 candidates.
            

            

              RRF_CONSTANT = 60
              RESULTS_PER_QUERY = 10
              MAX_RERANK_CANDIDATES = 20
            

          


          

            
Evidence reranking

            

              Claude ranks up to 20 fused candidates against the natural question,
              prioritizing direct doctrinal evidence over keyword overlap.
            

            

              rerank_and_select()
              doctrinal entailment
              Claude Sonnet 4.6 (default)
            

          


          

            
Grounding selection

            

              The validated ranking defines the eligible evidence. Final generation
              may cite one to three passages, but only from those selected sources.
            

            

              ranked_verse_ids[:2]
              validate_ranking()
              cited grounding
            

          


          

            
Grounded generation

            

              A final Bedrock call receives the question and selected verse context.
              If the selected evidence is weak, the app shows the closest verse
              as context instead of forcing a confident answer.
            

            

              build_context()
              generate_passage_message()
              faithfulness guardrails
              Claude Sonnet 4.6 (default)
            

          


          

            
Evaluation harness

            

              A 25-question regression set checks whether natural questions land on
              acceptable verses, so improvements can be tested instead of guessed.
            

            

              eval/rag_eval_cases.json
              evaluate_rag.py
              selected verse accuracy
            

          


          

            
Orchestration and API

            

              The final orchestration composes the grounded answer and exposes it
              through a validated FastAPI endpoint for the web experience and API consumers.
            

            

              answer_question()
              create_answer()
              trace evidence strength
            

          

        

      

      


      

        
          Build journey
How a 700-verse experiment became production RAG

          
        
        

          

            
The first version could retrieve a verse and generate a reflection. That was enough for a demo, but it was not enough to trust. The work since then has been about finding where the pipeline failed and replacing each weak link with something measurable.

          

          
    
            
  1.  The beginning
    One question in, one plausible verse out
    I started by indexing all 700 Bhagavad Gita verses with Amazon Titan embeddings and using Bedrock to turn a life situation into a response. The system worked, but semantic similarity alone often returned verses that sounded related without directly answering the question.
    2. 
            
  2.  The retrieval problem
    Broader search without losing precision
    I added four intent facets, deterministic Gita vocabulary, BM25 lexical search, and Reciprocal Rank Fusion. The pipeline now searches the original question and several controlled interpretations, then rewards verses that appear consistently across retrieval methods.
    3. 
            
  3.  The BGE phase
    Better ranking, with the wrong production cost
    BAAI/bge-reranker-v2-m3 was introduced as a local cross-encoder to compare the strongest candidates. Relevance improved, but model startup, image size, memory use, and latency made it a poor fit for this deployment. Reducing the number of comparisons helped, but did not solve the underlying trade-off.
    4. 
            
  4.  The replacement
    Ranking evidence instead of matching words
    BGE was removed and replaced with Claude evidence reranking. Claude now ranks six candidates by whether a verse directly supports the user’s situation, records evidence as strong, partial, or none, and limits final generation to the validated passages.
    5. 
            
  5.  The trust layer
    Make every answer inspectable
    The response now exposes cited verses, claim-level support, missing context, pipeline timings, generated queries, RRF candidates, and the final evidence ranking. When support is weak, the system says so and presents the closest passage instead of manufacturing certainty.
    6. 
            
  6.  Today
    A deployed and continuously evaluated system
    The production path runs from GitHub Actions to Docker, ECR, and EKS. After each successful deployment, the workflow executes a five-case smoke test followed by the full 25-case regression suite inside the running container. The result is a RAG system whose retrieval, grounding, latency, and failures can all be examined.
    700 versesHybrid retrievalEvidence-rankedCitation-grounded25-case regression
    7.