Hybrid Search

Find the context behind the question.

The words in a question may differ from the words in a memory. Combine keyword and similarity search to find useful context, then connect graph and resource retrieval where your application needs them.

Start free Read the docs

Search with context.

A question goes in.
Context comes back.

Context, connected.

The launch moved.
The reason didn’t get lost.

The team needed another week to finish the integration. The decision is in Monday’s conversation; the new date is in the launch brief.

Team conversation Launch brief

Give the next answer more to work with.

Bring relevant evidence into your agent’s response, with the source still attached.

Example application combining memory, graph, and resource retrieval.Explore each step above

More than a keyword match

One question. More places to find the answer.

Ask why a customer paused a project. The useful context could include their last conversation, a delivery procedure, and the original brief. Combine retrieval methods in your application to find relevant evidence across the memory it keeps.

Remember the conversation

Search past events and decisions in episodic memory. Recover what happened with the context attached to the experience.

Recognize the meaning

Similarity search can find related information without an exact phrase match. A question about a delayed launch can lead to a note about a postponed release.

Follow a relationship

Graph retrieval connects the question to people, projects, and related facts. Use relationships to add context around a match.

Return to the source

Search stored resources such as briefs and reports. Retrieve the original material your agent needs to support its response.

Keep exact details in reach

Names, product codes, and precise phrases matter. Keyword search complements similarity search when the exact wording is important.

Bring the results together

The engine offers weighted scoring and rank fusion to combine search results. Choose the retrieval strategy your application needs to produce a focused list.

Build a context-aware response

Bring relevant memories into a response.

Search the episodic memory index with a natural-language question. The HTTP endpoint returns matching record IDs and relevance scores, ready for your application to look up and use in its response.

See how ranked results come together

Hybrid Search · example· combining ranked resultsfuse

query

Why did Q3 revenue dip?

Episodicpast experiences

User asked why Q3 dipped0.91

Similar pattern in Q3 20240.74

Semanticsimilar meaning

Channel-mix shift drove the dip0.88

Paid vs organic mix vector0.71

Graphconnected facts

Revenue —driven_by→ Paid ch.0.83

Q3 —during→ mix-shift0.69

Resourcestored documents

Q3 board memo.pdf0.86

finance.query notes0.64

Fused rank

01User asked why Q3 dippedepisodic0.016

02Channel-mix shift drove the dipsemantic0.016

03Revenue —driven_by→ Paid ch.graph0.016

04Q3 board memo.pdfresource0.016

05Similar pattern in Q3 2024episodic0.016

06Paid vs organic mix vectorsemantic0.016

07Q3 —during→ mix-shiftgraph0.016

08finance.query notesresource0.016

Different searches. One result list.
Illustrative matches and scores

Search the episodic retrieval indexbash

# Use your configured Mind endpoint and an authorized token.
curl "$AKASHA_URL/v1/memory/retrieval/hybrid" \
  -H "Authorization: Bearer $AKASHA_TOKEN" \
  -H "x-akasha-capability: $AKASHA_CAPABILITY" \
  -H "Content-Type: application/json" \
  -d '{"text":"Why did Q3 revenue dip?","limit":5}'

# Response: {"rows":[{"id":"<record-id>","score":0.82}]}
# Example score; results depend on the stored information.

Technical details

A clear interface for richer retrieval.

The HTTP memory endpoint combines keyword and vector retrieval over the episodic index. Connect the graph and resource interfaces in your application when a question needs additional context.

Search request